taxonomy.json

{
  "version": "4.0",
  "extracted_from": "textbook",
  "extraction_model": "gemini-3.1-pro-preview",
  "extraction_date": "2026-03-23",
  "reviewed_by": "gemini-3.1-pro-preview + claude-opus-4.6",
  "concepts": [
    {
      "id": "3d-parallelism",
      "name": "3D Parallelism",
      "description": "The simultaneous application of data, tensor, and pipeline parallelism to successfully train massive frontier-scale machine learning models.",
      "prerequisites": [
        "pipeline-parallelism",
        "gradient-accumulation",
        "mini-batch-processing",
        "tensor-parallelism",
        "allreduce"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_training",
        "vol2_collective_communication",
        "vol2_distributed_training",
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "collective_communication",
        "distributed_training",
        "network_fabrics",
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "a-b-testing",
      "name": "A/B Testing",
      "description": "Comparing model variants in production with controlled traffic splitting.",
      "prerequisites": [
        "hardware-memory-hierarchy"
      ],
      "tracks": [
        "cloud",
        "mobile"
      ],
      "source_chapters": [
        "vol1_ml_ops"
      ],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "accelerator-bubbles",
      "name": "Accelerator Bubbles",
      "description": "Intervals of idle silicon where compute units sit at zero percent utilization waiting for data transfer or synchronization to complete.",
      "prerequisites": [
        "staged-system-pipeline"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "accelerator-spectrum",
      "name": "Accelerator Spectrum",
      "description": "The continuum of processing hardware ranging from highly programmable general-purpose CPUs to highly efficient, workload-specific custom ASICs.",
      "prerequisites": [
        "processing-elements"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "activation-checkpointing",
      "name": "Activation Checkpointing",
      "description": "A memory management strategy that trades increased computation time for reduced memory footprint by selectively recomputing intermediate activations during the backward pass.",
      "prerequisites": [
        "memory-bound-workload",
        "compute-bound-workload",
        "activation-memory",
        "reverse-mode-differentiation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks",
        "vol1_ml_systems",
        "vol1_training"
      ],
      "source_domains": [
        "frameworks",
        "ml_systems",
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "activation-logic-unit-cost",
      "name": "Activation Logic Unit Cost",
      "description": "The physical silicon area and transistor count required to implement a nonlinear activation function in hardware, favoring simple operators like ReLU over complex exponentials.",
      "prerequisites": [
        "activation-memory-footprint"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_computation"
      ],
      "source_domains": [
        "nn_computation"
      ],
      "question_count": 0
    },
    {
      "id": "activation-memory",
      "name": "Activation Memory",
      "description": "The storage required to cache forward-pass intermediate values for use during gradient computation in the backward pass.",
      "prerequisites": [
        "automatic-differentiation",
        "computational-graph"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "activation-memory-footprint",
      "name": "Activation Memory Footprint",
      "description": "The massive memory capacity required during training to store all intermediate layer outputs from the forward pass for reuse in gradient computation during backpropagation.",
      "prerequisites": [
        "mini-batch-processing",
        "activation-memory"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_computation"
      ],
      "source_domains": [
        "nn_computation"
      ],
      "question_count": 0
    },
    {
      "id": "activation-sparsity",
      "name": "Activation Sparsity",
      "description": "The property introduced by specific non-linearities that zero out negative inputs, reducing the computational memory footprint and energy usage by deactivating parts of the network dynamically.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_computation"
      ],
      "source_domains": [
        "nn_computation"
      ],
      "question_count": 0
    },
    {
      "id": "active-active-replication",
      "name": "Active-Active Replication",
      "description": "A serving redundancy strategy where all replicas actively process requests and load is distributed among them.",
      "prerequisites": [
        "stateless-serving"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "active-optical-cable",
      "name": "Active Optical Cable",
      "description": "Fiber cables with permanently attached transceivers used for medium-range connections that require power and add latency.",
      "prerequisites": [
        "forward-error-correction"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "adaptive-batch-size-scaling",
      "name": "Adaptive Batch Size Scaling",
      "description": "Adjusting global learning rates linearly as total batch size changes due to elastic worker scaling, keeping per-worker batch size fixed.",
      "prerequisites": [
        "elastic-training"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fleet_orchestration"
      ],
      "source_domains": [
        "fleet_orchestration"
      ],
      "question_count": 0
    },
    {
      "id": "adaptive-routing",
      "name": "Adaptive Routing",
      "description": "A dynamic routing mechanism that selects output ports based on real-time queue depth rather than static hashes.",
      "prerequisites": [
        "ecmp"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "adversarial-attacks",
      "name": "Adversarial Attacks",
      "description": "Deliberate attempts to manipulate model behavior and cause misclassification by adding imperceptible perturbations to inputs.",
      "prerequisites": [
        "offensive-machine-learning",
        "adversarial-inputs"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_robust_ai"
      ],
      "source_domains": [
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "adversarial-debiasing",
      "name": "Adversarial Debiasing",
      "description": "Training with an adversary that penalizes the model for learning group membership, preventing the encoding of proxy variables.",
      "prerequisites": [
        "stationarity-assumption"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr"
      ],
      "source_domains": [
        "responsible_engr"
      ],
      "question_count": 0
    },
    {
      "id": "adversarial-inputs",
      "name": "Adversarial Inputs",
      "description": "Carefully constructed input perturbations designed to cause a machine learning model to output incorrect or harmful predictions.",
      "prerequisites": [
        "model-memorization"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_responsible_ai"
      ],
      "source_domains": [
        "responsible_ai"
      ],
      "question_count": 0
    },
    {
      "id": "adversarial-training",
      "name": "Adversarial Training",
      "description": "A defense strategy that improves model robustness by augmenting the training dataset with adversarial examples.",
      "prerequisites": [
        "adversarial-attacks"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_robust_ai"
      ],
      "source_domains": [
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "ahead-of-time-optimization",
      "name": "Ahead-of-Time Optimization",
      "description": "Global graph optimizations like constant folding and dead code elimination performed by a framework compiler before runtime execution.",
      "prerequisites": [
        "static-computation-graphs"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "alignment-gap",
      "name": "Alignment Gap",
      "description": "A failure mode caused by misalignment between the technical proxy metric a model optimizes and its intended business purpose.",
      "prerequisites": [
        "stationarity-assumption"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr"
      ],
      "source_domains": [
        "responsible_engr"
      ],
      "question_count": 0
    },
    {
      "id": "allgather",
      "name": "AllGather",
      "description": "A collective communication pattern that collects different model portions so every device can reconstruct the full model state.",
      "prerequisites": [
        "bisection-bandwidth"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "allreduce",
      "name": "AllReduce",
      "description": "A collective communication primitive where all nodes share and aggregate their local arrays, leaving every node with the identical final reduced result.",
      "prerequisites": [
        "gradient-synchronization"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_collective_communication",
        "vol2_distributed_training",
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "collective_communication",
        "distributed_training",
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "alltoall-communication",
      "name": "AllToAll Communication",
      "description": "A dense network communication pattern required by Mixture-of-Experts models where every accelerator must simultaneously exchange distinct data with every other accelerator.",
      "prerequisites": [
        "embedding-table-sharding",
        "bisection-bandwidth"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_architectures",
        "vol2_collective_communication",
        "vol2_compute_infrastructure",
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "collective_communication",
        "compute_infrastructure",
        "network_fabrics",
        "nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "alpha-beta-performance-model",
      "name": "Alpha-Beta Performance Model",
      "description": "A linear network cost model that computes total data transfer time as a function of fixed startup latency and sustained per-byte bandwidth.",
      "prerequisites": [
        "latency"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_collective_communication",
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "collective_communication",
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "amdahl-s-law",
      "name": "Amdahl's Law",
      "description": "The mathematical principle defining the maximum possible speedup from hardware acceleration based on the non-parallelizable fraction of a workload.",
      "prerequisites": [
        "data-gravity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration",
        "vol1_ml_systems"
      ],
      "source_domains": [
        "hw_acceleration",
        "ml_systems"
      ],
      "question_count": 0
    },
    {
      "id": "amdahl-s-law-for-data",
      "name": "Amdahl's Law for Data",
      "description": "The mathematical limit on speedups achievable through parallelizing distributed data processing, constrained by operations requiring global coordination.",
      "prerequisites": [
        "batch-ingestion"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering"
      ],
      "source_domains": [
        "data_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "application-embedded-serving",
      "name": "Application-Embedded Serving",
      "description": "Running the model within the user's application process, optimizing for energy efficiency and responsiveness without a network hop.",
      "prerequisites": [
        "inference-server"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "approximate-model-theft",
      "name": "Approximate Model Theft",
      "description": "Training a surrogate model that mimics the original by systematically observing input-output behavior through an API.",
      "prerequisites": [
        "confidence-calibration"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "arithmetic-intensity",
      "name": "Arithmetic Intensity",
      "description": "The ratio of floating-point operations performed to bytes of memory traffic required during execution.",
      "prerequisites": [
        "computational-intensity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration",
        "vol1_ml_systems",
        "vol1_nn_architectures",
        "vol1_training",
        "vol2_compute_infrastructure",
        "vol2_performance_engineering",
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "compute_infrastructure",
        "hw_acceleration",
        "ml_systems",
        "nn_architectures",
        "performance_engineering",
        "sustainable_ai",
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "asynchronous-prefetching",
      "name": "Asynchronous Prefetching",
      "description": "Overlapping data loading and preprocessing with accelerator computation to prevent pipeline starvation.",
      "prerequisites": [
        "accelerator-bubbles"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "autograd-tape",
      "name": "Autograd Tape",
      "description": "A transient data structure built dynamically during eager execution that records the history of tensor operations to enable reverse-mode gradient computation.",
      "prerequisites": [
        "eager-execution",
        "automatic-differentiation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "automatic-differentiation",
      "name": "Automatic Differentiation",
      "description": "The systematic application of the chain rule in software to compute exact gradients for arbitrary compositions of operations.",
      "prerequisites": [
        "computational-graph"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "autoregressive-kv-cache",
      "name": "Autoregressive KV Cache",
      "description": "An inference memory buffer that stores previously computed key and value vectors to avoid redundant computation, shifting the system bottleneck entirely to memory bandwidth.",
      "prerequisites": [
        "bandwidth-bound-workload"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [
        "nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "backdoor-attacks",
      "name": "Backdoor Attacks",
      "description": "A type of data poisoning that embeds a hidden trigger causing the model to misclassify specific inputs while otherwise behaving normally.",
      "prerequisites": [
        "data-poisoning"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_robust_ai"
      ],
      "source_domains": [
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "backfill-scheduling",
      "name": "Backfill Scheduling",
      "description": "A utilization strategy that allows small, short jobs to run in fragmented resource gaps while large jobs wait for contiguous nodes.",
      "prerequisites": [
        "resource-fragmentation"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fleet_orchestration"
      ],
      "source_domains": [
        "fleet_orchestration"
      ],
      "question_count": 0
    },
    {
      "id": "bandwidth-bound-workload",
      "name": "Bandwidth-Bound Workload",
      "description": "A processing regime where execution time is limited by the speed of transferring data between memory and compute units.",
      "prerequisites": [
        "arithmetic-intensity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [
        "nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "bandwidth-distance-product",
      "name": "Bandwidth-Distance Product",
      "description": "A physical constraint dictating that a network link can provide either high data throughput or cover long physical distances, but not both cheaply.",
      "prerequisites": [
        "latency"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_collective_communication"
      ],
      "source_domains": [
        "collective_communication"
      ],
      "question_count": 0
    },
    {
      "id": "bandwidth-hog",
      "name": "Bandwidth Hog",
      "description": "A workload archetype that spends more time loading data than computing, constrained heavily by memory bandwidth.",
      "prerequisites": [
        "memory-bound-workload",
        "arithmetic-intensity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_systems"
      ],
      "source_domains": [
        "ml_systems"
      ],
      "question_count": 0
    },
    {
      "id": "bandwidth-taper",
      "name": "Bandwidth Taper",
      "description": "The physical constraint where data transfer speeds progressively decrease at each hierarchical level moving outward from the compute units.",
      "prerequisites": [
        "matrix-operations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "source_domains": [
        "hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "bare-metal-serving",
      "name": "Bare-Metal Serving",
      "description": "Executing models directly on microcontroller firmware without an operating system, optimizing for extreme memory constraints.",
      "prerequisites": [
        "tinyml"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "batch-ingestion",
      "name": "Batch Ingestion",
      "description": "Accumulating and processing data in scheduled, discrete groups to efficiently utilize computational resources.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering"
      ],
      "source_domains": [
        "data_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "batch-normalization",
      "name": "Batch Normalization",
      "description": "Normalizes activations within a mini-batch to stabilize training.",
      "prerequisites": [
        "dense-matrix-multiplication"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "batch-processing-trade-offs",
      "name": "Batch Processing Trade-offs",
      "description": "The system engineering balance where increasing the number of concurrent inputs improves parallel hardware utilization but scales the activation memory footprint linearly.",
      "prerequisites": [
        "activation-memory-footprint"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_computation"
      ],
      "source_domains": [
        "nn_computation"
      ],
      "question_count": 0
    },
    {
      "id": "batching-window",
      "name": "Batching Window",
      "description": "A configurable time period the scheduler waits to accumulate incoming requests into a single hardware batch.",
      "prerequisites": [
        "dynamic-batching"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "benchmark-contamination",
      "name": "Benchmark Contamination",
      "description": "A failure mode in large language models where the training corpus includes benchmark test data, inflating evaluation scores through memorization.",
      "prerequisites": [
        "model-benchmarking"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_benchmarking"
      ],
      "source_domains": [
        "benchmarking"
      ],
      "question_count": 0
    },
    {
      "id": "bin-packing-heuristics",
      "name": "Bin Packing Heuristics",
      "description": "Algorithms used to fit multi-dimensional job resource requests into fixed-capacity nodes to minimize wasted capacity.",
      "prerequisites": [
        "fleet-orchestration"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fleet_orchestration"
      ],
      "source_domains": [
        "fleet_orchestration"
      ],
      "question_count": 0
    },
    {
      "id": "binary-neural-networks",
      "name": "Binary Neural Networks",
      "description": "An extreme quantization approach representing weights and activations as single bits to enable sub-milliwatt inference on edge devices.",
      "prerequisites": [
        "model-quantization"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "bisection-bandwidth",
      "name": "Bisection Bandwidth",
      "description": "The minimum bandwidth available between two equal halves of a network, dictating global collective performance.",
      "prerequisites": [],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_distributed_training",
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "distributed_training",
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "block-wise-quantization",
      "name": "Block-wise Quantization",
      "description": "Grouping weights or activations to share scaling factors, isolating numerical outliers while preserving precision for the majority of values.",
      "prerequisites": [
        "precision-engineering"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "blue-green-deployment",
      "name": "Blue-Green Deployment",
      "description": "Instantaneously switching traffic between two identical production environments to isolate risk and enable fast rollbacks.",
      "prerequisites": [
        "sutva-violations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_ops_scale"
      ],
      "source_domains": [
        "ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "boundary-erosion",
      "name": "Boundary Erosion",
      "description": "The dissolution of system modularity where changes propagate unpredictably through implicit data dependencies rather than explicit code interfaces.",
      "prerequisites": [
        "stage-interface-contracts",
        "leaky-interfaces"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_ops"
      ],
      "source_domains": [
        "ml_ops"
      ],
      "question_count": 0
    },
    {
      "id": "bucket-fusion",
      "name": "Bucket Fusion",
      "description": "The practice of grouping multiple small gradient tensors into larger contiguous memory buffers before transmission to minimize network startup latency taxes.",
      "prerequisites": [
        "critical-message-size",
        "communication-computation-overlap"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_collective_communication"
      ],
      "source_domains": [
        "collective_communication"
      ],
      "question_count": 0
    },
    {
      "id": "bulk-synchronous-parallel",
      "name": "Bulk Synchronous Parallel",
      "description": "A distributed execution model where all nodes must complete their communication before any can proceed to the next compute step.",
      "prerequisites": [
        "straggler-mitigation",
        "gradient-synchronization"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "byzantine-failures",
      "name": "Byzantine Failures",
      "description": "Insidious failures where a component continues operating but produces incorrect results, violating the assumption of correct execution.",
      "prerequisites": [
        "fault-tolerance"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "byzantine-fault-tolerance",
      "name": "Byzantine Fault Tolerance",
      "description": "Robust aggregation mechanisms designed to maintain system reliability and global model convergence despite delayed, corrupted, or malicious device updates.",
      "prerequisites": [
        "federated-learning"
      ],
      "tracks": [
        "edge"
      ],
      "source_chapters": [
        "vol2_edge_intelligence"
      ],
      "source_domains": [
        "edge_intelligence"
      ],
      "question_count": 0
    },
    {
      "id": "cache-hierarchy",
      "name": "Cache Hierarchy",
      "description": "L1/L2/L3 cache levels with decreasing speed and increasing size.",
      "prerequisites": [
        "matrix-operations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "caching-allocators",
      "name": "Caching Allocators",
      "description": "A memory management abstraction that preemptively reserves large blocks of GPU memory to prevent fragmentation and avoid expensive system-level allocation bottlenecks.",
      "prerequisites": [
        "computational-graph"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "canary-deployment",
      "name": "Canary Deployment",
      "description": "A risk-mitigation strategy that routes a small percentage of live production traffic to a new model to validate performance stability.",
      "prerequisites": [
        "service-level-objective",
        "progressive-deployment"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_ops",
        "vol2_ops_scale"
      ],
      "source_domains": [
        "ml_ops",
        "ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "capacity-planning",
      "name": "Capacity Planning",
      "description": "Calculating required hardware resources by analyzing traffic patterns, latency targets, and model inference speeds.",
      "prerequisites": [
        "cost-per-inference",
        "service-level-objective"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "carbon-aware-scheduling",
      "name": "Carbon-Aware Scheduling",
      "description": "Scheduling intensive computational workloads during periods of abundant renewable energy to minimize environmental impact.",
      "prerequisites": [
        "operational-carbon"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr",
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "responsible_engr",
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "cascade-inference",
      "name": "Cascade Inference",
      "description": "An architectural pattern using a small edge model to filter inputs before conditionally escalating complex cases to the cloud.",
      "prerequisites": [
        "cmos-power-equation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "catastrophic-forgetting",
      "name": "Catastrophic Forgetting",
      "description": "A failure mode where an adapting model completely overwrites its previously learned general representations upon encountering distinct new local data.",
      "prerequisites": [
        "on-device-learning"
      ],
      "tracks": [
        "edge"
      ],
      "source_chapters": [
        "vol2_edge_intelligence"
      ],
      "source_domains": [
        "edge_intelligence"
      ],
      "question_count": 0
    },
    {
      "id": "centralized-checkpointing",
      "name": "Centralized Checkpointing",
      "description": "A simple distributed checkpointing approach where all workers send their state to a single coordinator process.",
      "prerequisites": [
        "checkpointing"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "certified-robustness",
      "name": "Certified Robustness",
      "description": "Mathematical guarantees ensuring a model's prediction will not change within a specified radius of input perturbation.",
      "prerequisites": [
        "adversarial-attacks"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_robust_ai"
      ],
      "source_domains": [
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "checkpoint-storms",
      "name": "Checkpoint Storms",
      "description": "Massive, simultaneous bursts of write operations from all cluster nodes to shared storage when saving model weights and optimizer states.",
      "prerequisites": [
        "sequential-throughput",
        "parallel-file-system"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "checkpointing",
      "name": "Checkpointing",
      "description": "Periodically writing the state of a model, optimizer, and training loop to durable storage to preserve progress.",
      "prerequisites": [
        "fault-tolerance"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "chiplet-architecture",
      "name": "Chiplet Architecture",
      "description": "A packaging design that combines multiple smaller semiconductor dies using high-bandwidth interconnects to bypass reticle limits and improve manufacturing yield.",
      "prerequisites": [
        "reticle-limit"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "circuit-breaker-pattern",
      "name": "Circuit Breaker Pattern",
      "description": "A reliability mechanism that stops calling a failing upstream service to prevent cascading resource exhaustion.",
      "prerequisites": [
        "batch-ingestion"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering"
      ],
      "source_domains": [
        "data_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "clean-label-poisoning",
      "name": "Clean-Label Poisoning",
      "description": "An attack that compromises models by injecting malicious samples into datasets without altering their labels to embed a hidden backdoor.",
      "prerequisites": [
        "data-poisoning"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_robust_ai"
      ],
      "source_domains": [
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "cmos-power-equation",
      "name": "CMOS Power Equation",
      "description": "The fundamental physics model separating chip power consumption into dynamic switching energy and static leakage current.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "co-packaged-optics",
      "name": "Co-Packaged Optics",
      "description": "A hardware design that integrates optical engines directly into the switch ASIC package to significantly reduce power consumption.",
      "prerequisites": [
        "bisection-bandwidth"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "cold-restart",
      "name": "Cold Restart",
      "description": "A recovery procedure where all processes are killed and the entire cluster state is reloaded from persistent storage.",
      "prerequisites": [
        "checkpointing"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "communication-bound",
      "name": "Communication-Bound",
      "description": "A system state where network synchronization saturates bandwidth, leaving accelerators idle while waiting for parameters or gradients.",
      "prerequisites": [
        "communication-computation-ratio"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_distributed_training"
      ],
      "source_domains": [
        "distributed_training"
      ],
      "question_count": 0
    },
    {
      "id": "communication-computation-overlap",
      "name": "Communication-Computation Overlap",
      "description": "A performance optimization that schedules network data transfers to occur simultaneously with independent mathematical computations, effectively hiding communication latency.",
      "prerequisites": [
        "cuda-streams"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_collective_communication",
        "vol2_network_fabrics",
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "collective_communication",
        "network_fabrics",
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "communication-computation-ratio",
      "name": "Communication-Computation Ratio",
      "description": "The ratio of time spent moving data across the network versus performing arithmetic operations on the accelerators.",
      "prerequisites": [],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_distributed_training"
      ],
      "source_domains": [
        "distributed_training"
      ],
      "question_count": 0
    },
    {
      "id": "compilation-continuum-principle",
      "name": "Compilation Continuum Principle",
      "description": "A quantitative decision framework evaluating when the runtime performance gains from JIT compilation justify its upfront development and compilation costs.",
      "prerequisites": [
        "just-in-time-compilation",
        "eager-execution"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "computation-placement",
      "name": "Computation Placement",
      "description": "The spatial assignment of specific operations to individual processing elements to maximize parallelism and minimize idle time.",
      "prerequisites": [
        "processing-elements"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "source_domains": [
        "hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "computational-graph",
      "name": "Computational Graph",
      "description": "A directed acyclic graph where nodes represent mathematical operations and edges represent data dependencies, used by frameworks to map execution.",
      "prerequisites": [
        "tensor-abstraction",
        "matrix-operations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "computational-intensity",
      "name": "Computational Intensity",
      "description": "The sustained petaFLOPS-scale throughput required by training workloads over long wall-clock durations.",
      "prerequisites": [
        "training-cost-asymmetry"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "compute-bound",
      "name": "Compute-Bound",
      "description": "A system state where accelerators spend the majority of their time multiplying matrices rather than waiting for data transfers.",
      "prerequisites": [
        "communication-computation-ratio"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_distributed_training"
      ],
      "source_domains": [
        "distributed_training"
      ],
      "question_count": 0
    },
    {
      "id": "compute-bound-workload",
      "name": "Compute-Bound Workload",
      "description": "A workload where execution speed is strictly limited by the processor's calculation throughput.",
      "prerequisites": [
        "arithmetic-intensity",
        "memory-wall"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks",
        "vol1_ml_systems",
        "vol1_nn_architectures",
        "vol1_training"
      ],
      "source_domains": [
        "frameworks",
        "ml_systems",
        "nn_architectures",
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "compute-express-link",
      "name": "Compute Express Link",
      "description": "An interconnect technology enabling memory pooling across nodes with load/store semantics instead of message passing.",
      "prerequisites": [
        "bisection-bandwidth"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "concept-drift",
      "name": "Concept Drift",
      "description": "A distribution shift where the underlying functional relationship between input features and target labels changes over time.",
      "prerequisites": [
        "covariate-shift",
        "distribution-shift",
        "silent-model-failures"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering",
        "vol1_ml_ops",
        "vol2_robust_ai"
      ],
      "source_domains": [
        "data_engineering",
        "ml_ops",
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "confidence-calibration",
      "name": "Confidence Calibration",
      "description": "Designing models to accurately reflect uncertainty or abstain from predictions when faced with adversarial or out-of-distribution inputs.",
      "prerequisites": [
        "secure-model-design"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "congestion-spreading",
      "name": "Congestion Spreading",
      "description": "A cascading failure in PFC-enabled networks where pause frames propagate upstream and freeze uninvolved links.",
      "prerequisites": [
        "priority-flow-control"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "constant-batch-size-scaling",
      "name": "Constant Batch Size Scaling",
      "description": "Maintaining optimization dynamics during elastic scaling by inversely adjusting per-worker batch sizes as the total worker count changes.",
      "prerequisites": [
        "elastic-training"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fleet_orchestration"
      ],
      "source_domains": [
        "fleet_orchestration"
      ],
      "question_count": 0
    },
    {
      "id": "constant-folding",
      "name": "Constant Folding",
      "description": "Pre-computing static parts of the computation graph at compile time to save inference cycles.",
      "prerequisites": [
        "application-embedded-serving"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "constraint-propagation",
      "name": "Constraint Propagation",
      "description": "The principle that late-stage constraint discoveries create exponentially escalating costs by forcing revisions across preceding stages.",
      "prerequisites": [
        "stage-interface-contracts"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_workflow"
      ],
      "source_domains": [
        "ml_workflow"
      ],
      "question_count": 0
    },
    {
      "id": "continuous-batching",
      "name": "Continuous Batching",
      "description": "An inference serving technique that dynamically schedules new requests at the iteration level to fill idle compute slots within an ongoing decode batch.",
      "prerequisites": [
        "static-batching",
        "model-flops-utilization"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure",
        "vol2_inference"
      ],
      "source_domains": [
        "compute_infrastructure",
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "continuous-deployment",
      "name": "Continuous Deployment",
      "description": "Operational practices enabling dynamic, continuous updates to production ML models based on automated monitoring and feedback loops.",
      "prerequisites": [
        "data-cascade"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_workflow"
      ],
      "source_domains": [
        "ml_workflow"
      ],
      "question_count": 0
    },
    {
      "id": "continuous-training-pipelines",
      "name": "Continuous Training Pipelines",
      "description": "Automated workflows that balance compute costs against accuracy by retraining models based on defined schedules or statistical drift triggers.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_ops"
      ],
      "source_domains": [
        "ml_ops"
      ],
      "question_count": 0
    },
    {
      "id": "contrastive-learning",
      "name": "Contrastive Learning",
      "description": "A self-supervised learning technique that encourages representations to capture invariant features across different views of the same data.",
      "prerequisites": [
        "self-supervised-learning"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_robust_ai"
      ],
      "source_domains": [
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "convolutional-neural-network",
      "name": "Convolutional Neural Network",
      "description": "Architecture using spatial convolutions for vision and signal processing.",
      "prerequisites": [
        "dense-matrix-multiplication"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "coreml",
      "name": "CoreML",
      "description": "Apple's on-device ML framework targeting Neural Engine, GPU, and CPU.",
      "prerequisites": [
        "on-device-learning"
      ],
      "tracks": [
        "mobile"
      ],
      "source_chapters": [
        "vol2_edge_intelligence"
      ],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "correction-cascades",
      "name": "Correction Cascades",
      "description": "A failure pattern where patching one ML component introduces statistical instabilities requiring endless downstream corrections.",
      "prerequisites": [
        "boundary-erosion"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_ops"
      ],
      "source_domains": [
        "ml_ops"
      ],
      "question_count": 0
    },
    {
      "id": "correlated-failures",
      "name": "Correlated Failures",
      "description": "Simultaneous failures of multiple components due to shared dependencies like power supplies or network switches.",
      "prerequisites": [
        "failure-rate"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "cost-per-inference",
      "name": "Cost Per Inference",
      "description": "The total economic cost of serving a single request, encompassing compute, memory reservation, and data transfer.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "counterfeit-hardware",
      "name": "Counterfeit Hardware",
      "description": "Unauthorized component reproductions that introduce reliability risks and potential hidden backdoors into machine learning supply chains.",
      "prerequisites": [
        "confidence-calibration"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "covariate-shift",
      "name": "Covariate Shift",
      "description": "A specific type of distribution shift where the input feature distribution changes while the relationship between inputs and outputs remains constant.",
      "prerequisites": [
        "distribution-shift",
        "kl-divergence"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_benchmarking",
        "vol1_data_engineering",
        "vol2_robust_ai"
      ],
      "source_domains": [
        "benchmarking",
        "data_engineering",
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "cpu-affinity",
      "name": "CPU Affinity",
      "description": "Pinning inference server threads to specific physical CPU cores to prevent context switching and maintain warm caches.",
      "prerequisites": [
        "latency-jitter"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "cpu-vectorization",
      "name": "CPU Vectorization",
      "description": "Mapping neural network operators directly to wide SIMD hardware instructions for accelerated CPU inference.",
      "prerequisites": [
        "application-embedded-serving"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "credit-assignment-problem",
      "name": "Credit Assignment Problem",
      "description": "The fundamental algorithmic challenge of calculating exactly how much each of the millions of individual weight parameters contributed to the final network prediction error.",
      "prerequisites": [
        "activation-memory-footprint"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_computation"
      ],
      "source_domains": [
        "nn_computation"
      ],
      "question_count": 0
    },
    {
      "id": "critical-message-size",
      "name": "Critical Message Size",
      "description": "The specific data payload threshold at which a network communication operation shifts from being bounded by latency to being bounded by bandwidth.",
      "prerequisites": [
        "alpha-beta-performance-model"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_collective_communication"
      ],
      "source_domains": [
        "collective_communication"
      ],
      "question_count": 0
    },
    {
      "id": "cuda-graphs",
      "name": "CUDA Graphs",
      "description": "A technique to record and replay a sequence of GPU operations to eliminate CPU-side kernel launch overhead.",
      "prerequisites": [
        "kernel-launch-overhead"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "cuda-streams",
      "name": "CUDA Streams",
      "description": "Ordered sequences of GPU operations that can execute concurrently with other streams to overlap compute and data movement.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "custom-asic",
      "name": "Custom ASIC",
      "description": "Specialized silicon chips designed to execute a specific computational workload with maximum energy efficiency at the cost of programmability.",
      "prerequisites": [
        "accelerator-spectrum"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "data-anonymization-layer",
      "name": "Data Anonymization Layer",
      "description": "An inference-path component applying privacy-preserving transformations like k-anonymity or differential privacy noise.",
      "prerequisites": [
        "model-memorization"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_responsible_ai"
      ],
      "source_domains": [
        "responsible_ai"
      ],
      "question_count": 0
    },
    {
      "id": "data-benchmarking",
      "name": "Data Benchmarking",
      "description": "The verification of whether a model generalizes to real-world data with its associated noise, bias, and distributional shift.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_benchmarking"
      ],
      "source_domains": [
        "benchmarking"
      ],
      "question_count": 0
    },
    {
      "id": "data-cascade",
      "name": "Data Cascade",
      "description": "The compounding technical debt and downstream system failures caused specifically by unmitigated data quality issues propagating forward.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering",
        "vol1_ml_workflow"
      ],
      "source_domains": [
        "data_engineering",
        "ml_workflow"
      ],
      "question_count": 0
    },
    {
      "id": "data-centric-ai",
      "name": "Data-Centric AI",
      "description": "A development paradigm that focuses on methodical dataset curation and enhancement to improve system performance rather than changing the model architecture.",
      "prerequisites": [
        "data-benchmarking"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_benchmarking"
      ],
      "source_domains": [
        "benchmarking"
      ],
      "question_count": 0
    },
    {
      "id": "data-gravity",
      "name": "Data Gravity",
      "description": "The constraint and cost of moving massive datasets across networks, calculated as volume divided by bandwidth.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering",
        "vol1_ml_systems"
      ],
      "source_domains": [
        "data_engineering",
        "ml_systems"
      ],
      "question_count": 0
    },
    {
      "id": "data-lineage",
      "name": "Data Lineage",
      "description": "The tracking of data provenance and metadata required for regulatory compliance and root cause analysis of model regressions.",
      "prerequisites": [
        "data-cascade"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_workflow"
      ],
      "source_domains": [
        "ml_workflow"
      ],
      "question_count": 0
    },
    {
      "id": "data-lineage-tracking",
      "name": "Data Lineage Tracking",
      "description": "Automated infrastructure recording the graph of relationships between raw data sources, feature transformations, and deployed model artifacts.",
      "prerequisites": [
        "stationarity-assumption"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr"
      ],
      "source_domains": [
        "responsible_engr"
      ],
      "question_count": 0
    },
    {
      "id": "data-locality-invariant",
      "name": "Data Locality Invariant",
      "description": "The rule stating data must remain local when transmission time exceeds total remote processing and network latency.",
      "prerequisites": [
        "data-gravity",
        "light-barrier"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_systems"
      ],
      "source_domains": [
        "ml_systems"
      ],
      "question_count": 0
    },
    {
      "id": "data-movement-energy-cost",
      "name": "Data Movement Energy Cost",
      "description": "The physical constraint where transferring model parameters and activations between memory and processing units consumes orders of magnitude more power than the actual mathematical operations.",
      "prerequisites": [
        "memory-wall"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_computation"
      ],
      "source_domains": [
        "nn_computation"
      ],
      "question_count": 0
    },
    {
      "id": "data-parallelism",
      "name": "Data Parallelism",
      "description": "Training strategy that replicates the model across devices and partitions the data batch.",
      "prerequisites": [
        "gradient-accumulation",
        "allreduce"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_training",
        "vol2_distributed_training"
      ],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "data-pipeline-equation",
      "name": "Data Pipeline Equation",
      "description": "A mathematical model used to calculate the aggregate storage bandwidth required to prevent accelerators from stalling.",
      "prerequisites": [
        "i-o-wall"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "data-poisoning",
      "name": "Data Poisoning",
      "description": "Deliberate corruption of the training process by injecting malicious samples to subvert learned representations and decision boundaries.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_robust_ai",
        "vol2_security_privacy"
      ],
      "source_domains": [
        "robust_ai",
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "data-sharding",
      "name": "Data Sharding",
      "description": "The practice of grouping thousands of individual data samples into large, contiguous files to enable high-throughput sequential reads.",
      "prerequisites": [
        "sequential-throughput"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "data-stall-ratio",
      "name": "Data Stall Ratio",
      "description": "The percentage of time during a training step that the accelerator sits idle waiting for data to arrive from storage.",
      "prerequisites": [
        "data-pipeline-equation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "data-striping",
      "name": "Data Striping",
      "description": "The distribution of segments of a single large file across multiple storage servers so clients can read from them simultaneously.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "dataflow-optimization",
      "name": "Dataflow Optimization",
      "description": "The strategic scheduling of data movement through hardware to maximize operand reuse and minimize expensive external memory accesses.",
      "prerequisites": [
        "computation-placement",
        "memory-allocation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "source_domains": [
        "hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "dataloader-choke-point",
      "name": "Dataloader Choke Point",
      "description": "An I/O bottleneck where CPU data decoding limits training throughput, leaving the GPU starved for data.",
      "prerequisites": [
        "feeding-tax"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering"
      ],
      "source_domains": [
        "data_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "dataset-compilation",
      "name": "Dataset Compilation",
      "description": "Reframing data engineering as a compiler pipeline that transforms raw observations into optimized, training-ready tensors.",
      "prerequisites": [
        "batch-ingestion"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering"
      ],
      "source_domains": [
        "data_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "datasheets-for-datasets",
      "name": "Datasheets for Datasets",
      "description": "Standardized artifacts documenting data provenance, collection methodology, and demographic composition of training data.",
      "prerequisites": [
        "model-memorization"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr",
        "vol2_responsible_ai"
      ],
      "source_domains": [
        "responsible_ai",
        "responsible_engr"
      ],
      "question_count": 0
    },
    {
      "id": "dcqcn",
      "name": "DCQCN",
      "description": "A reactive congestion control algorithm using Explicit Congestion Notification to prompt senders to reduce injection rates.",
      "prerequisites": [
        "roce"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "dead-letter-queues",
      "name": "Dead Letter Queues",
      "description": "Dedicated storage mechanisms for isolating data that repeatedly fails processing, preventing complete pipeline blockages.",
      "prerequisites": [
        "schema-validation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering"
      ],
      "source_domains": [
        "data_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "declarative-deployment-api",
      "name": "Declarative Deployment API",
      "description": "Abstracting container orchestration, resource allocation, and traffic mesh configurations into a unified deployment specification.",
      "prerequisites": [
        "sutva-violations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_ops_scale"
      ],
      "source_domains": [
        "ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "declarative-scheduling",
      "name": "Declarative Scheduling",
      "description": "A resource management model where users specify a desired operational state and continuous control loops reconcile the actual cluster state to match it.",
      "prerequisites": [
        "fleet-orchestration"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fleet_orchestration"
      ],
      "source_domains": [
        "fleet_orchestration"
      ],
      "question_count": 0
    },
    {
      "id": "degradation-equation",
      "name": "Degradation Equation",
      "description": "A formal model quantifying how fast a system's quality decays relative to the magnitude of distributional data shift.",
      "prerequisites": [
        "system-entropy"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_systems"
      ],
      "source_domains": [
        "ml_systems"
      ],
      "question_count": 0
    },
    {
      "id": "demographic-parity",
      "name": "Demographic Parity",
      "description": "A fairness metric requiring equal positive prediction rates across different demographic groups regardless of actual qualification.",
      "prerequisites": [
        "disaggregated-evaluation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr",
        "vol2_responsible_ai"
      ],
      "source_domains": [
        "responsible_ai",
        "responsible_engr"
      ],
      "question_count": 0
    },
    {
      "id": "dennard-scaling",
      "name": "Dennard Scaling",
      "description": "A classical physical law where the relationship between power and clock frequency was cubic, which forced the industry toward multi-core parallelism.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_systems",
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure",
        "ml_systems"
      ],
      "question_count": 0
    },
    {
      "id": "dense-connectivity-scaling",
      "name": "Dense Connectivity Scaling",
      "description": "An architectural pattern where full connections between layers cause the parameter count and resulting memory storage requirements to scale quadratically with layer width.",
      "prerequisites": [
        "activation-memory-footprint"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_computation"
      ],
      "source_domains": [
        "nn_computation"
      ],
      "question_count": 0
    },
    {
      "id": "dense-matrix-multiplication",
      "name": "Dense Matrix Multiplication",
      "description": "The foundational all-to-all computational pattern of MLPs that utilizes structured parallel multiply-accumulate hardware units.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [
        "nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "dependency-aware-deployment",
      "name": "Dependency-Aware Deployment",
      "description": "Coordinating model releases by ensuring upstream updates do not break downstream consumer models.",
      "prerequisites": [
        "model-registry"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_ops_scale"
      ],
      "source_domains": [
        "ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "depthwise-separable-convolution",
      "name": "Depthwise Separable Convolution",
      "description": "An architectural factorization that splits standard convolution into independent spatial and channel-wise operations to drastically reduce FLOPs for resource-constrained edge systems.",
      "prerequisites": [
        "arithmetic-intensity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_systems",
        "vol1_nn_architectures"
      ],
      "source_domains": [
        "ml_systems",
        "nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "differential-privacy",
      "name": "Differential Privacy",
      "description": "A framework providing mathematical guarantees about individual privacy by injecting calibrated noise during computation.",
      "prerequisites": [
        "federated-learning"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr",
        "vol2_edge_intelligence",
        "vol2_security_privacy"
      ],
      "source_domains": [
        "edge_intelligence",
        "responsible_engr",
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "direct-attach-copper",
      "name": "Direct Attach Copper",
      "description": "Passive copper cables offering the lowest latency and cost for short-range intra-rack connections.",
      "prerequisites": [
        "bisection-bandwidth"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "disaggregated-evaluation",
      "name": "Disaggregated Evaluation",
      "description": "Evaluating model metrics separately across demographic subgroups to reveal disparities hidden by aggregate performance.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr"
      ],
      "source_domains": [
        "responsible_engr"
      ],
      "question_count": 0
    },
    {
      "id": "disaggregated-serving",
      "name": "Disaggregated Serving",
      "description": "Separating the compute-bound prefill phase and memory-bound decode phase onto specialized, distinct hardware pools.",
      "prerequisites": [
        "time-per-output-token",
        "time-to-first-token"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "disparate-impact-ratio",
      "name": "Disparate Impact Ratio",
      "description": "A quantitative metric representing the ratio of favorable outcome rates between unprivileged and privileged groups.",
      "prerequisites": [
        "demographic-parity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_responsible_ai"
      ],
      "source_domains": [
        "responsible_ai"
      ],
      "question_count": 0
    },
    {
      "id": "dispatch-overhead-law",
      "name": "Dispatch Overhead Law",
      "description": "A principle showing that framework overhead dominates execution time when individual operations are small relative to the framework's launch latency.",
      "prerequisites": [
        "dispatch-tax"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "dispatch-tax",
      "name": "Dispatch Tax",
      "description": "The performance overhead incurred by the host-side interpreter when validating and dispatching operations to device-side silicon.",
      "prerequisites": [
        "eager-execution"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "distributed-checkpointing",
      "name": "Distributed Checkpointing",
      "description": "An approach where each worker writes its own portion of the checkpoint directly to shared storage, avoiding a central network bottleneck.",
      "prerequisites": [
        "checkpointing"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "distribution-shift",
      "name": "Distribution Shift",
      "description": "The phenomenon where the statistical distribution of data encountered in production differs from the training data distribution.",
      "prerequisites": [
        "stationarity-assumption",
        "data-benchmarking"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_benchmarking",
        "vol1_ml_ops",
        "vol1_ml_workflow",
        "vol1_responsible_engr",
        "vol2_responsible_ai",
        "vol2_robust_ai"
      ],
      "source_domains": [
        "benchmarking",
        "ml_ops",
        "ml_workflow",
        "responsible_ai",
        "responsible_engr",
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "double-descent-phenomenon",
      "name": "Double Descent Phenomenon",
      "description": "The counterintuitive scaling behavior where massive model overparameterization paradoxically improves generalization and reduces test error beyond traditional statistical limits.",
      "prerequisites": [
        "activation-memory-footprint"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_computation"
      ],
      "source_domains": [
        "nn_computation"
      ],
      "question_count": 0
    },
    {
      "id": "dragonfly-topology",
      "name": "Dragonfly Topology",
      "description": "A topology organizing switches into fully connected local groups linked by limited global optical connections.",
      "prerequisites": [
        "bisection-bandwidth"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "duty-cycle-accounting",
      "name": "Duty Cycle Accounting",
      "description": "A method for calculating average edge device power by weighting the energy of active inference bursts and deep sleep idle periods.",
      "prerequisites": [
        "cmos-power-equation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "dying-relu-problem",
      "name": "Dying ReLU Problem",
      "description": "A state where a significant percentage of a network's neurons permanently output zero and cease learning, resulting in wasted computational capacity and inefficient memory usage.",
      "prerequisites": [
        "activation-sparsity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_computation"
      ],
      "source_domains": [
        "nn_computation"
      ],
      "question_count": 0
    },
    {
      "id": "dynamic-attention-routing",
      "name": "Dynamic Attention Routing",
      "description": "A mechanism that computes mathematical relevance between input elements via queries and keys to generate data-dependent execution pathways.",
      "prerequisites": [
        "dense-matrix-multiplication",
        "autoregressive-kv-cache"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [
        "nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "dynamic-batching",
      "name": "Dynamic Batching",
      "description": "Accumulating requests up to either a maximum batch size or a maximum time window to balance throughput against strict latency targets.",
      "prerequisites": [
        "throughput",
        "latency",
        "static-batching",
        "inference-server"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving",
        "vol2_inference"
      ],
      "source_domains": [
        "inference",
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "dynamic-benchmarking",
      "name": "Dynamic Benchmarking",
      "description": "An evaluation approach that continuously updates test datasets to prevent saturation as model capabilities advance.",
      "prerequisites": [
        "data-benchmarking"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_benchmarking"
      ],
      "source_domains": [
        "benchmarking"
      ],
      "question_count": 0
    },
    {
      "id": "dynamic-inference",
      "name": "Dynamic Inference",
      "description": "Computing predictions on demand as user requests arrive, requiring strict latency management.",
      "prerequisites": [
        "latency",
        "inference-pipeline"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "dynamic-power",
      "name": "Dynamic Power",
      "description": "Energy consumed by switching transistors between states which scales quadratically with supply voltage.",
      "prerequisites": [
        "cmos-power-equation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "dynamic-voltage-frequency-scaling",
      "name": "Dynamic Voltage Frequency Scaling",
      "description": "A power management technique that adjusts a processor's clock speed and voltage in real time to maximize performance while staying within thermal limits.",
      "prerequisites": [
        "thermal-design-power"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "eager-execution",
      "name": "Eager Execution",
      "description": "A dynamic framework execution model where mathematical operations run immediately as encountered, maximizing debuggability but preventing global graph optimization.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "ecmp",
      "name": "ECMP",
      "description": "A static routing protocol that distributes traffic by hashing flow headers to fixed paths.",
      "prerequisites": [],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "efficiency-frontier",
      "name": "Efficiency Frontier",
      "description": "The Pareto-optimal curve balancing model quality against system throughput, latency, and cost.",
      "prerequisites": [
        "cuda-streams"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "ego-exclusion-design",
      "name": "Ego-Exclusion Design",
      "description": "Filtering out control group users with high treatment network exposure to recover unbiased causal effect estimates.",
      "prerequisites": [
        "sutva-violations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_ops_scale"
      ],
      "source_domains": [
        "ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "elastic-training",
      "name": "Elastic Training",
      "description": "Allowing a training job to dynamically resize and redistribute data when the number of available workers changes.",
      "prerequisites": [
        "fleet-orchestration",
        "checkpointing"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance",
        "vol2_fleet_orchestration"
      ],
      "source_domains": [
        "fault_tolerance",
        "fleet_orchestration"
      ],
      "question_count": 0
    },
    {
      "id": "elastic-weight-consolidation",
      "name": "Elastic Weight Consolidation",
      "description": "An algorithmic defense against catastrophic forgetting that applies mathematical penalties to prevent changes to weights critical for previously learned tasks.",
      "prerequisites": [
        "catastrophic-forgetting"
      ],
      "tracks": [
        "edge"
      ],
      "source_chapters": [
        "vol2_edge_intelligence"
      ],
      "source_domains": [
        "edge_intelligence"
      ],
      "question_count": 0
    },
    {
      "id": "elephant-flows",
      "name": "Elephant Flows",
      "description": "Massive, long-lasting data transfers that dominate ML training traffic, such as gradient AllReduce operations.",
      "prerequisites": [
        "bisection-bandwidth"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "elt-processing",
      "name": "ELT Processing",
      "description": "An architecture that loads raw data directly into storage and applies transformations within the target system to maximize experimentation flexibility.",
      "prerequisites": [
        "batch-ingestion"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering"
      ],
      "source_domains": [
        "data_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "embedding-table-sharding",
      "name": "Embedding Table Sharding",
      "description": "A distributed model parallelism strategy that partitions massive, memory-capacity-bound recommendation tables across multiple GPUs to overcome single-node limits.",
      "prerequisites": [
        "sparse-embedding-lookups",
        "model-parallelism"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [
        "nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "embodied-carbon",
      "name": "Embodied Carbon",
      "description": "Irreducible baseline greenhouse gas emissions generated during the manufacturing, material extraction, and supply chain logistics of hardware.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "energy-benchmarking",
      "name": "Energy Benchmarking",
      "description": "The measurement of power consumption and algorithmic energy efficiency in computing systems under ML workloads.",
      "prerequisites": [
        "system-benchmarking"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_benchmarking"
      ],
      "source_domains": [
        "benchmarking"
      ],
      "question_count": 0
    },
    {
      "id": "energy-movement-invariant",
      "name": "Energy-Movement Invariant",
      "description": "The fundamental physical constraint that moving data across networks consumes significantly more energy than computation.",
      "prerequisites": [
        "data-gravity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering"
      ],
      "source_domains": [
        "data_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "energy-per-byte",
      "name": "Energy Per Byte",
      "description": "The energy cost of moving data across the memory hierarchy which often dominates the cost of arithmetic operations.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "energy-per-operation",
      "name": "Energy Per Operation",
      "description": "The fundamental metric of computational efficiency typically measured in picojoules per multiply-accumulate operation.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "energy-roofline-model",
      "name": "Energy Roofline Model",
      "description": "An analytical framework visualizing whether an operation's energy consumption is bottlenecked by arithmetic computation or memory data movement.",
      "prerequisites": [
        "arithmetic-intensity",
        "energy-per-byte",
        "energy-per-operation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "equality-of-opportunity",
      "name": "Equality of Opportunity",
      "description": "A fairness constraint focusing solely on ensuring the true positive rate is equal across different demographic groups.",
      "prerequisites": [
        "equalized-odds"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr",
        "vol2_responsible_ai"
      ],
      "source_domains": [
        "responsible_ai",
        "responsible_engr"
      ],
      "question_count": 0
    },
    {
      "id": "equalized-odds",
      "name": "Equalized Odds",
      "description": "A fairness metric requiring both equal true positive rates and equal false positive rates across demographic groups.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr",
        "vol2_responsible_ai"
      ],
      "source_domains": [
        "responsible_ai",
        "responsible_engr"
      ],
      "question_count": 0
    },
    {
      "id": "erasure-coding",
      "name": "Erasure Coding",
      "description": "A mathematical method for achieving extreme data durability by splitting data into recoverable fragments with parity, avoiding full replication.",
      "prerequisites": [
        "object-storage-streaming"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "error-correcting-codes",
      "name": "Error Correcting Codes",
      "description": "Hardware-level mechanisms that add redundant bits to memory or transmissions to detect and correct transient bit flips.",
      "prerequisites": [
        "transient-failures"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "error-feedback",
      "name": "Error Feedback",
      "description": "A mathematical technique used alongside compression to accumulate and apply truncation errors in subsequent training steps, preventing model divergence.",
      "prerequisites": [
        "gradient-compression"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_collective_communication"
      ],
      "source_domains": [
        "collective_communication"
      ],
      "question_count": 0
    },
    {
      "id": "error-model",
      "name": "Error Model",
      "description": "A representation of how a localized fault propagates and alters the higher-level behavior of an ML system.",
      "prerequisites": [
        "fault-injection"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "event-driven-activation",
      "name": "Event-Driven Activation",
      "description": "An asynchronous computing paradigm that saves energy by processing only changing signals rather than continuously computing dense matrices.",
      "prerequisites": [
        "cmos-power-equation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "exact-model-theft",
      "name": "Exact Model Theft",
      "description": "Gaining direct access to and extracting a model's internal components, including serialized files and weights.",
      "prerequisites": [
        "confidence-calibration"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "execution-streams",
      "name": "Execution Streams",
      "description": "Independent execution queues on accelerators that enable the framework to overlap concurrent computation with asynchronous data transfers.",
      "prerequisites": [
        "pinned-memory"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "expected-calibration-error",
      "name": "Expected Calibration Error",
      "description": "A metric measuring the gap between a model's predicted confidence scores and its actual accuracy across different confidence bins.",
      "prerequisites": [
        "model-benchmarking"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_benchmarking"
      ],
      "source_domains": [
        "benchmarking"
      ],
      "question_count": 0
    },
    {
      "id": "experience-replay",
      "name": "Experience Replay",
      "description": "Maintaining a limited local storage buffer of historical data to interleave with new observations during training, mitigating catastrophic forgetting.",
      "prerequisites": [
        "catastrophic-forgetting"
      ],
      "tracks": [
        "edge"
      ],
      "source_chapters": [
        "vol2_edge_intelligence"
      ],
      "source_domains": [
        "edge_intelligence"
      ],
      "question_count": 0
    },
    {
      "id": "external-fragmentation",
      "name": "External Fragmentation",
      "description": "Unusable gaps of free memory created between active allocations as sequences of varying lengths complete and deallocate.",
      "prerequisites": [
        "pre-allocated-memory-management"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "fail-stop-failures",
      "name": "Fail-Stop Failures",
      "description": "Failures where a component ceases operation entirely and detectably, such as a node crash or network partition.",
      "prerequisites": [
        "fault-tolerance"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "failure-domains",
      "name": "Failure Domains",
      "description": "Hierarchical boundaries (e.g., node, rack, datacenter) used to isolate redundant components and prevent correlated failures.",
      "prerequisites": [
        "correlated-failures"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "failure-rate",
      "name": "Failure Rate",
      "description": "The expected frequency of component failures, typically measured in Mean Time Between Failures (MTBF) or Failures In Time (FIT).",
      "prerequisites": [
        "fault-tolerance"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "fair-share-scheduling",
      "name": "Fair-Share Scheduling",
      "description": "A policy algorithm that dynamically adjusts queue priorities based on a user's historical resource consumption to prevent cluster monopolization.",
      "prerequisites": [
        "fleet-orchestration"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fleet_orchestration"
      ],
      "source_domains": [
        "fleet_orchestration"
      ],
      "question_count": 0
    },
    {
      "id": "fairness-accuracy-pareto-frontier",
      "name": "Fairness-Accuracy Pareto Frontier",
      "description": "A curve representing the optimal configurations where improving algorithmic fairness necessarily degrades predictive accuracy.",
      "prerequisites": [
        "stationarity-assumption"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr"
      ],
      "source_domains": [
        "responsible_engr"
      ],
      "question_count": 0
    },
    {
      "id": "fairness-aware-training",
      "name": "Fairness-aware Training",
      "description": "In-processing methods that embed group fairness constraints directly into the optimization objective.",
      "prerequisites": [
        "model-memorization"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_responsible_ai"
      ],
      "source_domains": [
        "responsible_ai"
      ],
      "question_count": 0
    },
    {
      "id": "fairness-impossibility-law",
      "name": "Fairness Impossibility Law",
      "description": "The mathematical proof that Demographic Parity, Equalized Odds, and Calibration cannot be simultaneously satisfied when group base rates differ.",
      "prerequisites": [
        "equalized-odds",
        "demographic-parity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_responsible_ai"
      ],
      "source_domains": [
        "responsible_ai"
      ],
      "question_count": 0
    },
    {
      "id": "fat-tree-topology",
      "name": "Fat-Tree Topology",
      "description": "A hierarchical network topology that guarantees full non-blocking bisection bandwidth using leaf, spine, and core switches.",
      "prerequisites": [
        "bisection-bandwidth"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "fault-injection",
      "name": "Fault Injection",
      "description": "The engineering discipline of deliberately introducing faults into a system to empirically verify robustness mechanisms.",
      "prerequisites": [
        "fault-tolerance"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "fault-injection-attack",
      "name": "Fault Injection Attack",
      "description": "Applying precisely timed physical disturbances, like voltage drops or lasers, to hardware to induce computational errors during inference.",
      "prerequisites": [
        "physical-tampering"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "fault-model",
      "name": "Fault Model",
      "description": "A definition of how a physical hardware fault manifests itself in the system in terms of duration, location, and granularity.",
      "prerequisites": [
        "fault-injection"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "fault-tolerance",
      "name": "Fault Tolerance",
      "description": "Designing systems where failures are expected, detected, isolated, and recovered from automatically without losing progress.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "feature-parallel-batching",
      "name": "Feature-Parallel Batching",
      "description": "Batching requests by feature type rather than whole requests to optimize sparse embedding lookups across sharded servers.",
      "prerequisites": [
        "kv-cache"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "feature-store",
      "name": "Feature Store",
      "description": "Centralized infrastructure that computes features once and serves them identically to both batch training and real-time inference pipelines.",
      "prerequisites": [
        "training-serving-skew"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_ops"
      ],
      "source_domains": [
        "ml_ops"
      ],
      "question_count": 0
    },
    {
      "id": "federated-analytics",
      "name": "Federated Analytics",
      "description": "Collecting aggregated, privacy-preserving performance metrics and confidence summaries from device populations without centralizing raw inference telemetry.",
      "prerequisites": [
        "federated-learning",
        "secure-aggregation"
      ],
      "tracks": [
        "edge"
      ],
      "source_chapters": [
        "vol2_edge_intelligence"
      ],
      "source_domains": [
        "edge_intelligence"
      ],
      "question_count": 0
    },
    {
      "id": "federated-learning",
      "name": "Federated Learning",
      "description": "A distributed architecture that trains models across edge nodes without centralizing sensitive raw data to preserve privacy.",
      "prerequisites": [
        "on-device-learning"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_workflow",
        "vol1_responsible_engr",
        "vol2_edge_intelligence",
        "vol2_robust_ai",
        "vol2_security_privacy"
      ],
      "source_domains": [
        "edge_intelligence",
        "ml_workflow",
        "responsible_engr",
        "robust_ai",
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "feeding-tax",
      "name": "Feeding Tax",
      "description": "The wall-clock time lost and GPU efficiency reduced due to I/O wait from slow storage pipelines.",
      "prerequisites": [
        "data-gravity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering"
      ],
      "source_domains": [
        "data_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "flash-attention",
      "name": "Flash Attention",
      "description": "IO-aware exact attention algorithm that tiles computation to minimize HBM accesses.",
      "prerequisites": [
        "io-aware-algorithm-design"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol1_nn_architectures",
        "vol2_inference"
      ],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "flashattention",
      "name": "FlashAttention",
      "description": "A hardware-aware algorithm that reformulates self-attention into a tiled computation to avoid materializing the full attention matrix in off-chip memory.",
      "prerequisites": [
        "memory-bound-workload",
        "online-softmax",
        "tiling-algorithms",
        "operator-fusion"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training",
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "performance_engineering",
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "fleet-orchestration",
      "name": "Fleet Orchestration",
      "description": "The discipline of deciding which jobs run, where they run, and how resources are shared across a distributed compute cluster.",
      "prerequisites": [],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fleet_orchestration"
      ],
      "source_domains": [
        "fleet_orchestration"
      ],
      "question_count": 0
    },
    {
      "id": "forward-error-correction",
      "name": "Forward Error Correction",
      "description": "A physical layer technique to control transmission errors that adds irreducible latency to every network hop.",
      "prerequisites": [
        "pam4"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "forward-mode-differentiation",
      "name": "Forward Mode Differentiation",
      "description": "An automatic differentiation technique that computes derivatives alongside the original computation using dual numbers, scaling efficiently with outputs but poorly with inputs.",
      "prerequisites": [
        "automatic-differentiation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "gang-scheduling",
      "name": "Gang Scheduling",
      "description": "An atomic, all-or-nothing allocation mechanism ensuring all distributed workers start simultaneously to prevent deadlocks.",
      "prerequisites": [
        "fleet-orchestration"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fleet_orchestration"
      ],
      "source_domains": [
        "fleet_orchestration"
      ],
      "question_count": 0
    },
    {
      "id": "generality-tax",
      "name": "Generality Tax",
      "description": "The energy and compute overhead incurred by CPUs to support flexible instruction sequences, which is unnecessary for uniform neural network workloads.",
      "prerequisites": [
        "accelerator-spectrum"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "go-back-n",
      "name": "Go-Back-N",
      "description": "A retransmission mechanism where a single dropped packet forces the sender to retransmit all subsequent packets.",
      "prerequisites": [],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "gpu-direct-storage",
      "name": "GPU Direct Storage",
      "description": "A hardware bypass technology establishing a direct DMA mapping to transfer data straight from NVMe drives to GPU memory, bypassing the CPU completely.",
      "prerequisites": [
        "high-bandwidth-memory",
        "local-nvme-warm-cache"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "gpu-memory-hierarchy",
      "name": "GPU Memory Hierarchy",
      "description": "The physical tiered storage architecture of an accelerator, spanning fast registers to large but slow high bandwidth memory.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "gpu-preprocessing",
      "name": "GPU Preprocessing",
      "description": "Executing data transformation steps on the accelerator to eliminate expensive CPU-to-GPU memory transfers.",
      "prerequisites": [
        "preprocessing-bottleneck"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "gpudirect-rdma",
      "name": "GPUDirect RDMA",
      "description": "An advanced hardware configuration that allows a network interface card to transfer data directly to and from GPU high-bandwidth memory over the PCIe bus without CPU involvement.",
      "prerequisites": [
        "remote-direct-memory-access"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_collective_communication",
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "collective_communication",
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "graceful-degradation",
      "name": "Graceful Degradation",
      "description": "Ensuring a system maintains core functionality with predictable performance reduction when components fail or operate under stress.",
      "prerequisites": [
        "adversarial-attacks"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_robust_ai"
      ],
      "source_domains": [
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "gradient-accumulation",
      "name": "Gradient Accumulation",
      "description": "Simulating larger effective batch sizes by computing and summing gradients across sequential micro-batches before updating parameters.",
      "prerequisites": [
        "mini-batch-processing",
        "activation-memory"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training",
        "vol2_distributed_training"
      ],
      "source_domains": [
        "distributed_training",
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "gradient-checkpointing",
      "name": "Gradient Checkpointing",
      "description": "Discarding intermediate activations during the forward pass and recomputing them on-demand during the backward pass to drastically reduce peak memory usage.",
      "prerequisites": [
        "on-device-learning"
      ],
      "tracks": [
        "edge"
      ],
      "source_chapters": [
        "vol2_edge_intelligence"
      ],
      "source_domains": [
        "edge_intelligence"
      ],
      "question_count": 0
    },
    {
      "id": "gradient-compression",
      "name": "Gradient Compression",
      "description": "Reducing the size of the communication payload by quantizing or sparsifying gradient tensors before transmitting them across the network.",
      "prerequisites": [
        "gradient-synchronization"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_collective_communication"
      ],
      "source_domains": [
        "collective_communication"
      ],
      "question_count": 0
    },
    {
      "id": "gradient-inversion-attack",
      "name": "Gradient Inversion Attack",
      "description": "An attack that reconstructs raw original training data directly from the gradient updates transmitted by client devices.",
      "prerequisites": [
        "federated-learning"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "gradient-quantization",
      "name": "Gradient Quantization",
      "description": "Reducing the numerical precision of calculated parameter updates to drastically lower the network bandwidth required for transmission.",
      "prerequisites": [
        "federated-learning"
      ],
      "tracks": [
        "edge"
      ],
      "source_chapters": [
        "vol2_edge_intelligence"
      ],
      "source_domains": [
        "edge_intelligence"
      ],
      "question_count": 0
    },
    {
      "id": "gradient-synchronization",
      "name": "Gradient Synchronization",
      "description": "The mathematical necessity of ensuring all distributed worker GPUs apply identical weight updates to maintain a consistent global model state.",
      "prerequisites": [],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_collective_communication"
      ],
      "source_domains": [
        "collective_communication"
      ],
      "question_count": 0
    },
    {
      "id": "graph-breaks",
      "name": "Graph Breaks",
      "description": "Costly interruptions in a compiled execution region caused by unsupported dynamic operations, forcing a fallback to slower eager Python execution.",
      "prerequisites": [
        "just-in-time-compilation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "graph-cluster-randomization",
      "name": "Graph Cluster Randomization",
      "description": "Mitigating network effect bias by assigning experimental treatments to dense user communities instead of individuals.",
      "prerequisites": [
        "sutva-violations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_ops_scale"
      ],
      "source_domains": [
        "ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "graph-compilation",
      "name": "Graph Compilation",
      "description": "Automated translation of high-level model definitions into optimized, hardware-aware machine instructions and fused kernels.",
      "prerequisites": [
        "operator-fusion"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "graph-scripting",
      "name": "Graph Scripting",
      "description": "A JIT compilation method that statically analyzes source code to preserve all control flow branches in the compiled intermediate representation.",
      "prerequisites": [
        "just-in-time-compilation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "graph-tracing",
      "name": "Graph Tracing",
      "description": "A JIT compilation method that records a single execution path to build a static graph, offering high fidelity but silently missing unexecuted control flow branches.",
      "prerequisites": [
        "just-in-time-compilation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "group-distributionally-robust-optimization",
      "name": "Group Distributionally Robust Optimization",
      "description": "A training technique that minimizes the loss of the worst-case group to force the model to learn features that generalize across contexts.",
      "prerequisites": [
        "spurious-correlations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_robust_ai"
      ],
      "source_domains": [
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "grouped-query-attention",
      "name": "Grouped Query Attention",
      "description": "An architectural design where multiple query heads share a single key-value head, fundamentally reducing the required KV cache size.",
      "prerequisites": [
        "kv-cache"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "grpc-protocol",
      "name": "gRPC Protocol",
      "description": "An HTTP/2-based interface using binary Protobuf serialization for highly efficient, low-latency internal communication.",
      "prerequisites": [
        "zero-copy-serialization"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "hardware-flops-utilization",
      "name": "Hardware FLOPs Utilization",
      "description": "The ratio of total executed arithmetic operations to the hardware's peak theoretical compute limit.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "hardware-memory-hierarchy",
      "name": "Hardware Memory Hierarchy",
      "description": "The multi-tiered memory architecture spanning L2 cache, high-bandwidth memory (HBM), and system RAM that dictates inference latency.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_ops"
      ],
      "source_domains": [
        "ml_ops"
      ],
      "question_count": 0
    },
    {
      "id": "hardware-power-counters",
      "name": "Hardware Power Counters",
      "description": "Dedicated on-chip registers that provide ground-truth microsecond-resolution measurements of hardware energy consumption.",
      "prerequisites": [
        "cmos-power-equation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "hardware-precision-scaling",
      "name": "Hardware Precision Scaling",
      "description": "A hardware mechanism that automatically monitors tensor magnitude distributions and dynamically toggles between numerical precisions to maximize throughput.",
      "prerequisites": [
        "tensor-cores"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "hardware-redundancy",
      "name": "Hardware Redundancy",
      "description": "Using component duplication and voting circuits, like Double Modular Redundancy, to detect and mask faults during computation.",
      "prerequisites": [
        "fault-tolerance"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "hardware-software-co-design",
      "name": "Hardware-Software Co-design",
      "description": "The design principle where algorithms are tailored to hardware capabilities and hardware is built around specific algorithmic requirements.",
      "prerequisites": [
        "matrix-operations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "source_domains": [
        "hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "hardware-trojans",
      "name": "Hardware Trojans",
      "description": "Malicious, dormant circuits embedded during chip fabrication that trigger under specific conditions to disrupt computation or leak outputs.",
      "prerequisites": [
        "physical-tampering"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "hbm",
      "name": "HBM",
      "description": "High Bandwidth Memory — stacked DRAM providing TB/s bandwidth for GPUs.",
      "prerequisites": [
        "gpu-memory-hierarchy"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_hw_acceleration",
        "vol2_compute_infrastructure"
      ],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "hedged-requests",
      "name": "Hedged Requests",
      "description": "A latency-mitigation technique where a redundant read request is sent to a different storage replica if the primary request times out.",
      "prerequisites": [
        "tail-latency"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "hierarchical-allreduce",
      "name": "Hierarchical AllReduce",
      "description": "A multi-level communication strategy that applies different collective algorithms to optimize traffic across distinct hardware topologies, such as using ring within a node and trees across nodes.",
      "prerequisites": [
        "ring-allreduce",
        "allreduce"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_collective_communication"
      ],
      "source_domains": [
        "collective_communication"
      ],
      "question_count": 0
    },
    {
      "id": "hierarchical-processing",
      "name": "Hierarchical Processing",
      "description": "A hybrid architecture pattern where data flows upward and is progressively processed across tiny, edge, and cloud tiers.",
      "prerequisites": [
        "tinyml"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_systems"
      ],
      "source_domains": [
        "ml_systems"
      ],
      "question_count": 0
    },
    {
      "id": "high-bandwidth-memory",
      "name": "High Bandwidth Memory",
      "description": "A memory architecture that vertically stacks DRAM dies connected by through-silicon vias to provide massive data bandwidth on the processor's package.",
      "prerequisites": [
        "gpu-memory-hierarchy",
        "ml-storage-hierarchy",
        "memory-wall"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure",
        "vol2_data_storage",
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "compute_infrastructure",
        "data_storage",
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "hold-and-wait-deadlock",
      "name": "Hold-and-Wait Deadlock",
      "description": "A systemic failure where multiple jobs hold partial resource allocations while waiting for more, preventing any from making progress.",
      "prerequisites": [
        "gang-scheduling"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fleet_orchestration"
      ],
      "source_domains": [
        "fleet_orchestration"
      ],
      "question_count": 0
    },
    {
      "id": "homomorphic-encryption",
      "name": "Homomorphic Encryption",
      "description": "A technique allowing machine learning inference or training to be computed directly on encrypted inputs without decrypting them.",
      "prerequisites": [
        "confidence-calibration"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "host-dram-staging",
      "name": "Host DRAM Staging",
      "description": "The use of main system memory as a prefetch buffer and workspace for decoding and augmenting data before transfer to the GPU.",
      "prerequisites": [
        "high-bandwidth-memory"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "host-networking-bypass",
      "name": "Host Networking Bypass",
      "description": "Configuring containerized workloads to bypass virtualized network interfaces for direct RDMA access, eliminating microservice latency overheads.",
      "prerequisites": [
        "declarative-scheduling"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fleet_orchestration"
      ],
      "source_domains": [
        "fleet_orchestration"
      ],
      "question_count": 0
    },
    {
      "id": "hpcc",
      "name": "HPCC",
      "description": "A proactive congestion control mechanism utilizing In-Network Telemetry metadata to precisely adjust transmission rates.",
      "prerequisites": [
        "bisection-bandwidth"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "i-o-jitter",
      "name": "I/O Jitter",
      "description": "Unpredictable variations in storage read latency caused by network contention or background hardware maintenance.",
      "prerequisites": [
        "tail-latency"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "i-o-wall",
      "name": "I/O Wall",
      "description": "The performance bottleneck that occurs when storage throughput cannot deliver training data as fast as the accelerators consume it.",
      "prerequisites": [
        "sequential-throughput"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "idempotent-transformations",
      "name": "Idempotent Transformations",
      "description": "Data processing operations that consistently yield identical outputs given identical inputs, regardless of how many times they execute.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering"
      ],
      "source_domains": [
        "data_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "im2col-transformation",
      "name": "im2col Transformation",
      "description": "An algorithmic technique that reshapes convolutional operations into standard matrix multiplications for efficient hardware execution.",
      "prerequisites": [
        "dense-matrix-multiplication",
        "matrix-operations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration",
        "vol1_nn_architectures"
      ],
      "source_domains": [
        "hw_acceleration",
        "nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "immutable-audit-trails",
      "name": "Immutable Audit Trails",
      "description": "Append-only storage systems that securely record every data access event for regulatory compliance and operational accountability.",
      "prerequisites": [
        "stationarity-assumption"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr"
      ],
      "source_domains": [
        "responsible_engr"
      ],
      "question_count": 0
    },
    {
      "id": "imperative-scheduling",
      "name": "Imperative Scheduling",
      "description": "A resource management paradigm where users specify exact cluster resources required and the system allocates them directly.",
      "prerequisites": [
        "fleet-orchestration"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fleet_orchestration"
      ],
      "source_domains": [
        "fleet_orchestration"
      ],
      "question_count": 0
    },
    {
      "id": "incast",
      "name": "Incast",
      "description": "A network congestion event occurring when thousands of nodes simultaneously transmit data to the same target switch port.",
      "prerequisites": [
        "bulk-synchronous-parallel"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "inductive-bias",
      "name": "Inductive Bias",
      "description": "A structural assumption encoded into a neural architecture to restrict its hypothesis space and improve learning efficiency for specific data types.",
      "prerequisites": [
        "dense-matrix-multiplication"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [
        "nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "inference-pipeline",
      "name": "Inference Pipeline",
      "description": "The complete execution path of a request, encompassing data preprocessing, neural network computation, and postprocessing.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "inference-server",
      "name": "Inference Server",
      "description": "A high-performance software architecture that schedules requests, manages memory, and bridges irregular user traffic with batch-oriented hardware.",
      "prerequisites": [
        "throughput",
        "latency"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "infiniband",
      "name": "InfiniBand",
      "description": "A purpose-built HPC network protocol featuring hardware-level credit-based flow control for inherently lossless delivery.",
      "prerequisites": [
        "remote-direct-memory-access"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "influence-functions",
      "name": "Influence Functions",
      "description": "A technique that approximates the effect of removing a single training data point on the model's loss for a specific test prediction.",
      "prerequisites": [
        "adversarial-attacks"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_robust_ai"
      ],
      "source_domains": [
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "input-stationary",
      "name": "Input Stationary",
      "description": "A dataflow strategy that keeps input activations fixed in local memory while streaming weights and partial sums through the processing elements.",
      "prerequisites": [
        "dataflow-optimization"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "source_domains": [
        "hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "integration-bottleneck",
      "name": "Integration Bottleneck",
      "description": "The primary performance limit in modern AI caused by the latency and energy cost of moving massive data to highly parallel compute units.",
      "prerequisites": [
        "memory-wall"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "source_domains": [
        "hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "interleaving-experiments",
      "name": "Interleaving Experiments",
      "description": "Evaluating model variants by simultaneously presenting a combined ranked list of predictions to a single user.",
      "prerequisites": [
        "sutva-violations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_ops_scale"
      ],
      "source_domains": [
        "ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "interrupt-shielding",
      "name": "Interrupt Shielding",
      "description": "Routing network and storage hardware interrupts to non-inference cores to avoid preempting GPU commands.",
      "prerequisites": [
        "latency-jitter"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "intersectional-fairness",
      "name": "Intersectional Fairness",
      "description": "The evaluation of model performance across overlapping identity attributes to prevent compounded biases.",
      "prerequisites": [
        "demographic-parity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_responsible_ai"
      ],
      "source_domains": [
        "responsible_ai"
      ],
      "question_count": 0
    },
    {
      "id": "invariance-testing",
      "name": "Invariance Testing",
      "description": "A testing strategy verifying that a model's predictions remain unchanged when irrelevant protected attributes are altered.",
      "prerequisites": [
        "stationarity-assumption"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr"
      ],
      "source_domains": [
        "responsible_engr"
      ],
      "question_count": 0
    },
    {
      "id": "io-aware-algorithm-design",
      "name": "IO-Aware Algorithm Design",
      "description": "Structuring algorithms to minimize data movement across memory hierarchies rather than just minimizing raw FLOP counts.",
      "prerequisites": [
        "memory-bound-workload",
        "arithmetic-intensity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "iron-law-of-ml-systems",
      "name": "Iron Law of ML Systems",
      "description": "An equation decomposing total ML execution latency into data movement, computation, and fixed overhead costs.",
      "prerequisites": [
        "power-wall",
        "memory-wall"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_systems"
      ],
      "source_domains": [
        "ml_systems"
      ],
      "question_count": 0
    },
    {
      "id": "iron-law-of-training",
      "name": "Iron Law of Training",
      "description": "A framework mapping training time performance to total operations, peak hardware throughput, and system utilization.",
      "prerequisites": [
        "computational-intensity",
        "accelerator-bubbles"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "iteration-tax",
      "name": "Iteration Tax",
      "description": "The compounding time and resource cost of slow iteration cycles caused by discovering constraint violations late in development.",
      "prerequisites": [
        "constraint-propagation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_workflow"
      ],
      "source_domains": [
        "ml_workflow"
      ],
      "question_count": 0
    },
    {
      "id": "jensen-shannon-divergence",
      "name": "Jensen-Shannon Divergence",
      "description": "A symmetric, bounded variant of Kullback-Leibler divergence used for practical drift detection thresholds.",
      "prerequisites": [
        "kl-divergence"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_robust_ai"
      ],
      "source_domains": [
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "just-in-time-compilation",
      "name": "Just-In-Time Compilation",
      "description": "A hybrid execution strategy that captures dynamic runtime computation to produce an optimized intermediate representation for subsequent executions.",
      "prerequisites": [
        "eager-execution",
        "computational-graph"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "kernel-fusion",
      "name": "Kernel Fusion",
      "description": "The compiler optimization technique of combining multiple consecutive memory-bound operations into a single hardware kernel to eliminate intermediate memory traffic.",
      "prerequisites": [
        "memory-bound-workload",
        "memory-boundedness",
        "computational-graph"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks",
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure",
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "kernel-launch-overhead",
      "name": "Kernel Launch Overhead",
      "description": "The latency incurred by the CPU preparing and dispatching commands to the GPU execution queue.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "kl-divergence",
      "name": "KL Divergence",
      "description": "An information-theoretic measure of distance between continuous data distributions used to detect model data drift.",
      "prerequisites": [
        "distribution-shift"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering",
        "vol2_robust_ai"
      ],
      "source_domains": [
        "data_engineering",
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "knowledge-distillation",
      "name": "Knowledge Distillation",
      "description": "An efficiency technique where a smaller, low-power student model is trained to approximate the outputs of a massive teacher model.",
      "prerequisites": [
        "cmos-power-equation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "kv-cache",
      "name": "KV Cache",
      "description": "Memory used to store intermediate key and value tensors during Large Language Model generation to avoid recomputing context.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving",
        "vol2_inference"
      ],
      "source_domains": [
        "inference",
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "kv-cache-bottleneck",
      "name": "KV-Cache Bottleneck",
      "description": "A memory capacity limitation in LLM serving where storing attention keys and values per sequence restricts maximum batch size.",
      "prerequisites": [
        "hardware-memory-hierarchy"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_ops"
      ],
      "source_domains": [
        "ml_ops"
      ],
      "question_count": 0
    },
    {
      "id": "kv-cache-checkpointing",
      "name": "KV Cache Checkpointing",
      "description": "Periodically saving the conversational key-value cache of an LLM session to avoid full regeneration upon replica failure.",
      "prerequisites": [
        "stateful-serving"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "kv-cache-quantization",
      "name": "KV Cache Quantization",
      "description": "Compressing the stored attention activations into lower precision formats like INT8 or INT4 to reduce memory footprint.",
      "prerequisites": [
        "kv-cache"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "label-quality-drift",
      "name": "Label Quality Drift",
      "description": "The gradual degradation in the reliability and consistency of ground truth annotations, independent of underlying data distributions.",
      "prerequisites": [
        "concept-drift"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering"
      ],
      "source_domains": [
        "data_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "label-shift",
      "name": "Label Shift",
      "description": "A distribution shift where the output label frequency changes while the relationship between labels and features remains constant.",
      "prerequisites": [
        "covariate-shift",
        "distribution-shift"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering",
        "vol2_robust_ai"
      ],
      "source_domains": [
        "data_engineering",
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "latency",
      "name": "Latency",
      "description": "The time required to process an individual request and return a response.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "latency-budget",
      "name": "Latency Budget",
      "description": "The explicit allocation of allowable processing time across preprocessing, inference, and postprocessing pipeline stages.",
      "prerequisites": [
        "service-level-objective",
        "inference-pipeline"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "latency-jitter",
      "name": "Latency Jitter",
      "description": "Unpredictable variations in request processing time caused by operating system resource contention.",
      "prerequisites": [
        "latency",
        "tail-latency"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "law-of-distributed-efficiency",
      "name": "Law of Distributed Efficiency",
      "description": "A principle stating that the ratio of communication time to computation time is determined almost entirely by the network fabric.",
      "prerequisites": [
        "bisection-bandwidth"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "layer-normalization",
      "name": "Layer Normalization",
      "description": "Normalizes activations across features within a single sample.",
      "prerequisites": [
        "dense-matrix-multiplication"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "layered-defense",
      "name": "Layered Defense",
      "description": "A security architecture employing multiple independent defensive mechanisms across data, model, runtime, and hardware layers.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "leaky-interfaces",
      "name": "Leaky Interfaces",
      "description": "Exposed communication or debug channels that unintentionally reveal sensitive information or accept unverified inputs.",
      "prerequisites": [
        "confidence-calibration"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "learnability-gap",
      "name": "Learnability Gap",
      "description": "The practical distinction between what a network can theoretically represent versus what it can successfully optimize given finite data and compute.",
      "prerequisites": [
        "dense-matrix-multiplication"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [
        "nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "lifecycle-carbon-accounting",
      "name": "Lifecycle Carbon Accounting",
      "description": "A comprehensive assessment framework combining operational emissions and amortized embodied carbon across a system's entire lifespan.",
      "prerequisites": [
        "operational-carbon",
        "embodied-carbon"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "light-barrier",
      "name": "Light Barrier",
      "description": "The absolute minimum latency floor dictated by the speed of light over network distances.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_systems"
      ],
      "source_domains": [
        "ml_systems"
      ],
      "question_count": 0
    },
    {
      "id": "load-balancer",
      "name": "Load Balancer",
      "description": "An infrastructure component that distributes requests across model replicas, monitors health, and manages traffic during rollouts.",
      "prerequisites": [
        "application-embedded-serving"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "load-balancing",
      "name": "Load Balancing",
      "description": "Distributing work evenly across resources to maximize utilization.",
      "prerequisites": [
        "fleet-orchestration"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fleet_orchestration"
      ],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "local-nvme-warm-cache",
      "name": "Local NVMe Warm Cache",
      "description": "The use of high-speed, node-local solid-state drives to stage data fetched from distributed storage, avoiding repeated network reads.",
      "prerequisites": [
        "data-sharding"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "locality-aware-scheduling",
      "name": "Locality-Aware Scheduling",
      "description": "The orchestration strategy of placing compute workloads on specific nodes that already have the required data cached locally.",
      "prerequisites": [
        "local-nvme-warm-cache"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "low-rank-adapters",
      "name": "Low-Rank Adapters",
      "description": "Inserting small trainable modules into a frozen pretrained model to achieve deep representation adaptation with minimal memory overhead.",
      "prerequisites": [
        "on-device-learning"
      ],
      "tracks": [
        "edge"
      ],
      "source_chapters": [
        "vol2_edge_intelligence"
      ],
      "source_domains": [
        "edge_intelligence"
      ],
      "question_count": 0
    },
    {
      "id": "m-g-c-k-queue-model",
      "name": "M/G/c/K Queue Model",
      "description": "A queuing theory model used to analyze GPU inference systems with stochastic arrival rates and general service time distributions.",
      "prerequisites": [
        "kv-cache"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "machine-unlearning",
      "name": "Machine Unlearning",
      "description": "The algorithmic process of removing the statistical influence of specific training data records from model weights.",
      "prerequisites": [
        "model-memorization"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_responsible_ai"
      ],
      "source_domains": [
        "responsible_ai"
      ],
      "question_count": 0
    },
    {
      "id": "matrix-multiplication-workloads",
      "name": "Matrix Multiplication Workloads",
      "description": "The core linear algebra operations that dominate computational time and determine if a neural network layer is limited by compute capabilities or memory bandwidth.",
      "prerequisites": [
        "multiply-accumulate-operations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_computation"
      ],
      "source_domains": [
        "nn_computation"
      ],
      "question_count": 0
    },
    {
      "id": "matrix-operations",
      "name": "Matrix Operations",
      "description": "Structured, multi-dimensional mathematical transformations that dominate the execution time of neural networks.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "source_domains": [
        "hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "maximum-mean-discrepancy",
      "name": "Maximum Mean Discrepancy",
      "description": "A statistical distance metric used to quantify distribution shift by measuring differences between training and deployment data distributions.",
      "prerequisites": [
        "distribution-shift"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_robust_ai"
      ],
      "source_domains": [
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "membership-inference-attack",
      "name": "Membership Inference Attack",
      "description": "Determining whether a specific individual's data record was included in a model's training dataset.",
      "prerequisites": [
        "model-memorization",
        "differential-privacy"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr",
        "vol2_responsible_ai",
        "vol2_security_privacy"
      ],
      "source_domains": [
        "responsible_ai",
        "responsible_engr",
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "memory-allocation",
      "name": "Memory Allocation",
      "description": "The temporal and spatial distribution of parameters and activations across the hardware's memory hierarchy.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "source_domains": [
        "hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "memory-bandwidth",
      "name": "Memory Bandwidth",
      "description": "Rate of data transfer between memory and compute (GB/s or TB/s).",
      "prerequisites": [
        "activation-memory-footprint"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration",
        "vol1_nn_computation"
      ],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "memory-bound-workload",
      "name": "Memory-Bound Workload",
      "description": "A workload where execution speed is strictly limited by memory bandwidth rather than processor calculation speed.",
      "prerequisites": [
        "arithmetic-intensity",
        "memory-wall"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks",
        "vol1_ml_systems",
        "vol1_training"
      ],
      "source_domains": [
        "frameworks",
        "ml_systems",
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "memory-boundedness",
      "name": "Memory Boundedness",
      "description": "A condition where workload performance is strictly bottlenecked by data transfer speeds from memory rather than the system's compute capacity.",
      "prerequisites": [
        "memory-wall"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "memory-capacity-bound-workload",
      "name": "Memory-Capacity-Bound Workload",
      "description": "A system state where execution is limited by the raw physical storage required for massive parameter tables, typical of recommendation systems.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [
        "nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "memory-locking",
      "name": "Memory Locking",
      "description": "Pinning model weights and caches in physical RAM to prevent the operating system from swapping them to disk.",
      "prerequisites": [
        "latency-jitter"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "memory-planning",
      "name": "Memory Planning",
      "description": "Pre-calculating exact memory addresses for tensors during graph compilation to eliminate dynamic allocation overhead.",
      "prerequisites": [
        "application-embedded-serving"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "memory-pressure",
      "name": "Memory Pressure",
      "description": "The combined storage requirements of model weights, optimizer states, and forward-pass activations that often exceed single-accelerator capacity.",
      "prerequisites": [
        "optimizer-memory-overhead",
        "activation-memory"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "memory-wall",
      "name": "Memory Wall",
      "description": "The growing architectural performance gap between processor computational speed and memory fetch bandwidth.",
      "prerequisites": [
        "gpu-memory-hierarchy",
        "matrix-multiplication-workloads"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks",
        "vol1_hw_acceleration",
        "vol1_ml_systems",
        "vol1_nn_computation",
        "vol2_compute_infrastructure",
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "compute_infrastructure",
        "frameworks",
        "hw_acceleration",
        "ml_systems",
        "nn_computation",
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "metadata-overhead",
      "name": "Metadata Overhead",
      "description": "The administrative computational delay introduced by file system operations like opening and tracking individual files.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "metric-cardinality",
      "name": "Metric Cardinality",
      "description": "The combinatorial explosion of unique monitoring labels that dominates the computational and storage costs of system observability.",
      "prerequisites": [
        "hardware-memory-hierarchy"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_ops"
      ],
      "source_domains": [
        "ml_ops"
      ],
      "question_count": 0
    },
    {
      "id": "mice-flows",
      "name": "Mice Flows",
      "description": "A vast number of small, independent, and asynchronous data flows typical of traditional web service traffic.",
      "prerequisites": [
        "bisection-bandwidth"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "mini-batch-processing",
      "name": "Mini-batch Processing",
      "description": "Grouping training examples to convert memory-bound vector operations into compute-bound matrix multiplications for better hardware utilization.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "mixed-precision-computing",
      "name": "Mixed-Precision Computing",
      "description": "The strategic use of different numerical formats throughout an execution pipeline to optimize speed and memory bandwidth while preserving accuracy.",
      "prerequisites": [
        "matrix-operations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "source_domains": [
        "hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "mixed-precision-training",
      "name": "Mixed-Precision Training",
      "description": "A technique combining reduced-precision arithmetic for bandwidth efficiency with high-precision storage to maintain numerical stability during gradient updates.",
      "prerequisites": [
        "compute-bound-workload",
        "memory-pressure",
        "reverse-mode-differentiation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks",
        "vol1_training",
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure",
        "frameworks",
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "mixture-of-experts",
      "name": "Mixture of Experts",
      "description": "An architectural optimization routing inputs to specialized sub-networks to decouple total parameter count from per-token compute cost.",
      "prerequisites": [
        "cuda-streams"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "ml-storage-hierarchy",
      "name": "ML Storage Hierarchy",
      "description": "A multi-tier storage architecture designed to bridge the massive gap between slow, high-capacity object storage and fast, low-capacity accelerator memory.",
      "prerequisites": [
        "i-o-wall"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "mlperf",
      "name": "MLPerf",
      "description": "Industry-standard benchmark suite for ML training and inference performance.",
      "prerequisites": [
        "data-benchmarking"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_benchmarking"
      ],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "model-benchmarking",
      "name": "Model Benchmarking",
      "description": "The validation of whether optimization and compression techniques preserve model quality across the full input distribution.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_benchmarking"
      ],
      "source_domains": [
        "benchmarking"
      ],
      "question_count": 0
    },
    {
      "id": "model-cards",
      "name": "Model Cards",
      "description": "Standardized documentation capturing technical details, intended uses, and disaggregated performance metrics for deployed models.",
      "prerequisites": [
        "disaggregated-evaluation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr",
        "vol2_responsible_ai"
      ],
      "source_domains": [
        "responsible_ai",
        "responsible_engr"
      ],
      "question_count": 0
    },
    {
      "id": "model-cold-start-latency",
      "name": "Model Cold-Start Latency",
      "description": "The time required to read and transfer a massive multi-gigabyte model from storage into HBM during scale-up or recovery.",
      "prerequisites": [
        "sequential-throughput",
        "high-bandwidth-memory"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "model-ensembles",
      "name": "Model Ensembles",
      "description": "Maintaining multiple models specialized for different conditions and dynamically selecting or weighting them to adapt to environmental shifts.",
      "prerequisites": [
        "distribution-shift"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_robust_ai"
      ],
      "source_domains": [
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "model-flops-utilization",
      "name": "Model FLOPS Utilization",
      "description": "The ratio of observed mathematical throughput during model execution compared to the theoretical peak compute performance of the hardware.",
      "prerequisites": [
        "roofline-model",
        "hardware-flops-utilization"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure",
        "vol2_distributed_training",
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "compute_infrastructure",
        "distributed_training",
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "model-format-optimization",
      "name": "Model Format Optimization",
      "description": "Converting trained models into specialized, hardware-efficient interchange representations like ONNX or TensorRT to accelerate inference.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_ops"
      ],
      "source_domains": [
        "ml_ops"
      ],
      "question_count": 0
    },
    {
      "id": "model-inversion-attack",
      "name": "Model Inversion Attack",
      "description": "Querying a model through an interface to reverse-engineer and infer private details about the original training data.",
      "prerequisites": [
        "confidence-calibration"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "model-memorization",
      "name": "Model Memorization",
      "description": "The tendency of models to encode specific details from training data, creating privacy vulnerabilities queryable at inference.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_responsible_ai"
      ],
      "source_domains": [
        "responsible_ai"
      ],
      "question_count": 0
    },
    {
      "id": "model-parallelism",
      "name": "Model Parallelism",
      "description": "A scaling strategy that partitions a single model's layers or operations across multiple accelerators when it exceeds single-device memory limits.",
      "prerequisites": [
        "memory-pressure"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "model-pruning",
      "name": "Model Pruning",
      "description": "A compression technique that improves energy efficiency by systematically eliminating redundant weights and computational paths from a neural network.",
      "prerequisites": [
        "cmos-power-equation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "model-quantization",
      "name": "Model Quantization",
      "description": "Reducing memory bandwidth requirements and increasing inference throughput by mapping floating-point model weights to lower-precision integers.",
      "prerequisites": [
        "memory-bound-workload",
        "energy-per-byte",
        "memory-boundedness",
        "dynamic-power",
        "model-format-optimization"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_ops",
        "vol1_ml_systems",
        "vol2_compute_infrastructure",
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "compute_infrastructure",
        "ml_ops",
        "ml_systems",
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "model-registry",
      "name": "Model Registry",
      "description": "A centralized, version-controlled repository managing trained model artifacts, evaluation metrics, and deployment state.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_ops",
        "vol2_ops_scale"
      ],
      "source_domains": [
        "ml_ops",
        "ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "model-sharding",
      "name": "Model Sharding",
      "description": "Distributing model parameters across multiple accelerators when they exceed the memory capacity of a single device.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "model-watermarking",
      "name": "Model Watermarking",
      "description": "Embedding verifiable, hidden ownership signatures directly into a model's parameters or output behavior to protect intellectual property.",
      "prerequisites": [
        "secure-model-design"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "multi-instance-gpu-mig",
      "name": "Multi-Instance GPU (MIG)",
      "description": "A technology that partitions a single physical GPU into multiple hardware-isolated instances with dedicated compute and memory.",
      "prerequisites": [
        "resource-fragmentation"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fleet_orchestration"
      ],
      "source_domains": [
        "fleet_orchestration"
      ],
      "question_count": 0
    },
    {
      "id": "multi-region-deployment",
      "name": "Multi-Region Deployment",
      "description": "Synchronizing model versions, traffic routing, and rollback protocols across geographically distributed inference infrastructure.",
      "prerequisites": [
        "canary-deployment"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_ops_scale"
      ],
      "source_domains": [
        "ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "multi-scale-feedback",
      "name": "Multi-Scale Feedback",
      "description": "The orchestration of monitoring loops across multiple timescales to enable both rapid operational correction and long-term strategic adaptation.",
      "prerequisites": [
        "data-cascade"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_workflow"
      ],
      "source_domains": [
        "ml_workflow"
      ],
      "question_count": 0
    },
    {
      "id": "multi-worker-data-loading",
      "name": "Multi-Worker Data Loading",
      "description": "The spawning of multiple independent CPU processes to parallelize I/O reads, decoding, and data augmentation.",
      "prerequisites": [
        "prefetch-buffer-depth"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "multicalibration",
      "name": "Multicalibration",
      "description": "An advanced strategy ensuring model predictions remain calibrated across a wide range of intersecting subgroups.",
      "prerequisites": [
        "intersectional-fairness"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_responsible_ai"
      ],
      "source_domains": [
        "responsible_ai"
      ],
      "question_count": 0
    },
    {
      "id": "multiple-testing-correction",
      "name": "Multiple Testing Correction",
      "description": "Applying statistical adjustments like False Discovery Rate control to prevent false positive deployments across concurrent experiments.",
      "prerequisites": [
        "sutva-violations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_ops_scale"
      ],
      "source_domains": [
        "ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "multiply-accumulate-operations",
      "name": "Multiply-Accumulate Operations",
      "description": "The foundational hardware-level arithmetic instruction that computes the weighted sums of neural inputs and defines the theoretical maximum throughput of an AI accelerator.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_computation"
      ],
      "source_domains": [
        "nn_computation"
      ],
      "question_count": 0
    },
    {
      "id": "networked-serving",
      "name": "Networked Serving",
      "description": "Deploying the model as a standalone microservice accessed over a network, typically optimizing for throughput via batching.",
      "prerequisites": [
        "application-embedded-serving"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "neural-processing-unit",
      "name": "Neural Processing Unit",
      "description": "A specialized hardware accelerator integrated into mobile and edge devices to execute ML tasks with high energy efficiency.",
      "prerequisites": [
        "data-gravity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_systems"
      ],
      "source_domains": [
        "ml_systems"
      ],
      "question_count": 0
    },
    {
      "id": "non-iid-data-distributions",
      "name": "Non-IID Data Distributions",
      "description": "Highly fragmented, non-uniform data environments across different edge devices that complicate distributed gradient convergence assumptions.",
      "prerequisites": [
        "federated-learning"
      ],
      "tracks": [
        "edge"
      ],
      "source_chapters": [
        "vol2_edge_intelligence"
      ],
      "source_domains": [
        "edge_intelligence"
      ],
      "question_count": 0
    },
    {
      "id": "numa-aware-allocation",
      "name": "NUMA-Aware Allocation",
      "description": "The practice of pinning memory allocations and data loader threads to the same physical CPU socket as the GPU to avoid interconnect bandwidth penalties.",
      "prerequisites": [
        "host-dram-staging"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "numa-aware-inference",
      "name": "NUMA-Aware Inference",
      "description": "Allocating memory and threads on the same physical CPU socket to avoid cross-socket communication delays.",
      "prerequisites": [
        "cpu-affinity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "nvlink",
      "name": "NVLink",
      "description": "High-bandwidth GPU-to-GPU interconnect (900 GB/s bidirectional on NVLink 4.0).",
      "prerequisites": [
        "bisection-bandwidth"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_hw_acceleration",
        "vol2_network_fabrics"
      ],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "object-storage-streaming",
      "name": "Object Storage Streaming",
      "description": "The process of issuing concurrent HTTP byte-range requests to fetch specific data chunks directly from highly durable blob storage.",
      "prerequisites": [
        "data-sharding"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "offensive-machine-learning",
      "name": "Offensive Machine Learning",
      "description": "The weaponization of machine learning models to automate, accelerate, or enhance cyberattack operations against other systems.",
      "prerequisites": [
        "confidence-calibration"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "on-device-learning",
      "name": "On-Device Learning",
      "description": "Training and adapting machine learning models directly on edge devices to preserve privacy and operate within latency and bandwidth limits.",
      "prerequisites": [],
      "tracks": [
        "edge"
      ],
      "source_chapters": [
        "vol2_edge_intelligence"
      ],
      "source_domains": [
        "edge_intelligence"
      ],
      "question_count": 0
    },
    {
      "id": "online-learning",
      "name": "Online Learning",
      "description": "A continuous adaptation strategy allowing models to update on new data streams to mitigate concept drift.",
      "prerequisites": [
        "concept-drift"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_robust_ai"
      ],
      "source_domains": [
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "online-softmax",
      "name": "Online Softmax",
      "description": "A mathematical trick maintaining running statistics to compute exact softmax scores incrementally without full matrix materialization.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "onnx-runtime",
      "name": "ONNX Runtime",
      "description": "Cross-platform inference engine supporting multiple hardware backends.",
      "prerequisites": [
        "computational-graph"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "operational-carbon",
      "name": "Operational Carbon",
      "description": "Greenhouse gas emissions resulting from electricity consumption during active model training and inference phases.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "operator-fusion",
      "name": "Operator Fusion",
      "description": "The optimization technique of combining multiple adjacent operations into a single kernel execution to eliminate intermediate memory round-trips.",
      "prerequisites": [
        "gpu-memory-hierarchy",
        "dataflow-optimization",
        "memory-wall"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration",
        "vol1_ml_systems",
        "vol1_model_serving",
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "hw_acceleration",
        "ml_systems",
        "model_serving",
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "opportunistic-training-scheduling",
      "name": "Opportunistic Training Scheduling",
      "description": "Executing local model updates exclusively when edge devices are idle, charging, and thermally stable to prevent degradation of the user experience.",
      "prerequisites": [
        "on-device-learning"
      ],
      "tracks": [
        "edge"
      ],
      "source_chapters": [
        "vol2_edge_intelligence"
      ],
      "source_domains": [
        "edge_intelligence"
      ],
      "question_count": 0
    },
    {
      "id": "optimizer-memory-overhead",
      "name": "Optimizer Memory Overhead",
      "description": "The additional memory required to store mathematical state like momentum and moving averages for parameter updates.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "output-stationary",
      "name": "Output Stationary",
      "description": "A dataflow strategy that keeps intermediate partial sums fixed in local memory to eliminate redundant memory write operations.",
      "prerequisites": [
        "dataflow-optimization"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "source_domains": [
        "hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "paged-kv-cache-management",
      "name": "Paged KV Cache Management",
      "description": "Allocating Key-Value cache memory in dynamic, fixed-size blocks to eliminate fragmentation and enable larger effective batch sizes.",
      "prerequisites": [
        "cuda-streams"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "paged-memory-management",
      "name": "Paged Memory Management",
      "description": "Allocating memory dynamically in fixed-size blocks to eliminate fragmentation and improve maximum concurrent request capacity.",
      "prerequisites": [
        "pre-allocated-memory-management"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "pagedattention-mechanism",
      "name": "PagedAttention Mechanism",
      "description": "An attention implementation that stores the KV cache in non-contiguous memory pages to completely eliminate fragmentation waste.",
      "prerequisites": [
        "kv-cache",
        "paged-memory-management"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "pam4",
      "name": "PAM4",
      "description": "A signal encoding method that packs more information into each transition by using four distinct voltage levels.",
      "prerequisites": [],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "parallel-file-system",
      "name": "Parallel File System",
      "description": "A distributed storage architecture that separates metadata management from data object servers to provide massive aggregate bandwidth.",
      "prerequisites": [
        "data-striping",
        "ml-storage-hierarchy"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "pareto-frontier",
      "name": "Pareto Frontier",
      "description": "A trade-off curve identifying models that cannot improve one efficiency metric without degrading another.",
      "prerequisites": [
        "model-benchmarking"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_benchmarking"
      ],
      "source_domains": [
        "benchmarking"
      ],
      "question_count": 0
    },
    {
      "id": "pcie",
      "name": "PCIe",
      "description": "Standard host-to-device interconnect (128 GB/s bidirectional for Gen5 x16).",
      "prerequisites": [
        "accelerator-spectrum"
      ],
      "tracks": [
        "cloud",
        "edge"
      ],
      "source_chapters": [
        "vol1_hw_acceleration",
        "vol2_compute_infrastructure"
      ],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "personalization-layers",
      "name": "Personalization Layers",
      "description": "Partitioning a network into a frozen shared backbone and a lightweight, device-specific final layer that is updated locally to save compute.",
      "prerequisites": [
        "on-device-learning"
      ],
      "tracks": [
        "edge"
      ],
      "source_chapters": [
        "vol2_edge_intelligence"
      ],
      "source_domains": [
        "edge_intelligence"
      ],
      "question_count": 0
    },
    {
      "id": "physical-tampering",
      "name": "Physical Tampering",
      "description": "Direct, unauthorized manipulation of physical computing hardware to undermine the integrity of machine learning systems.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "pinned-memory",
      "name": "Pinned Memory",
      "description": "Page-locked CPU memory that enables the accelerator to perform Direct Memory Access (DMA) transfers without interrupting the host processor.",
      "prerequisites": [
        "tensor-abstraction"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "pipeline-bubble",
      "name": "Pipeline Bubble",
      "description": "The idle compute time GPUs spend waiting for activations or gradients to propagate through the stages of a pipeline.",
      "prerequisites": [
        "pipeline-parallelism"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_distributed_training"
      ],
      "source_domains": [
        "distributed_training"
      ],
      "question_count": 0
    },
    {
      "id": "pipeline-parallelism",
      "name": "Pipeline Parallelism",
      "description": "Overlapping the execution of partitioned model segments by splitting batches into sequential microbatches to reduce idle time.",
      "prerequisites": [
        "accelerator-bubbles",
        "model-parallelism"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training",
        "vol2_distributed_training",
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "distributed_training",
        "network_fabrics",
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "pipelined-execution",
      "name": "Pipelined Execution",
      "description": "A hardware technique that overlaps data movement with computation to maximize throughput.",
      "prerequisites": [
        "data-gravity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_systems"
      ],
      "source_domains": [
        "ml_systems"
      ],
      "question_count": 0
    },
    {
      "id": "population-stability-index",
      "name": "Population Stability Index",
      "description": "A statistical metric used to quantify how much a categorical or binned feature distribution has shifted over time.",
      "prerequisites": [
        "covariate-shift",
        "distribution-shift",
        "schema-validation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering",
        "vol1_ml_ops",
        "vol2_robust_ai"
      ],
      "source_domains": [
        "data_engineering",
        "ml_ops",
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "post-hoc-explainability",
      "name": "Post-Hoc Explainability",
      "description": "Methods like SHAP or LIME that approximate feature importance scores for individual predictions without altering opaque model architectures.",
      "prerequisites": [
        "system-explainability"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr",
        "vol2_responsible_ai"
      ],
      "source_domains": [
        "responsible_ai",
        "responsible_engr"
      ],
      "question_count": 0
    },
    {
      "id": "post-training-quantization",
      "name": "Post-Training Quantization",
      "description": "Converting a fully trained model to lower precision without a full retraining process using a small calibration dataset.",
      "prerequisites": [
        "precision-engineering"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "power-analysis-attack",
      "name": "Power Analysis Attack",
      "description": "A specific side-channel attack that measures power consumption patterns to infer internal system states or recover secret data.",
      "prerequisites": [
        "side-channel-attack"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "power-density-ceiling",
      "name": "Power Density Ceiling",
      "description": "The absolute physical limit on datacenter computational capacity dictated by available electrical infrastructure and cooling.",
      "prerequisites": [
        "cmos-power-equation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "power-usage-effectiveness",
      "name": "Power Usage Effectiveness",
      "description": "A facility-level metric measuring the ratio of total data center power to the power consumed strictly by computing equipment.",
      "prerequisites": [
        "cmos-power-equation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "power-wall",
      "name": "Power Wall",
      "description": "The thermodynamic limit on how much computation can occur in a given volume before heat causes hardware throttling.",
      "prerequisites": [
        "dennard-scaling"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_systems"
      ],
      "source_domains": [
        "ml_systems"
      ],
      "question_count": 0
    },
    {
      "id": "pre-allocated-memory-management",
      "name": "Pre-allocated Memory Management",
      "description": "Reserving a fixed maximum memory budget per request at admission time, often leading to significant wasted capacity.",
      "prerequisites": [
        "memory-allocation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "precision-engineering",
      "name": "Precision Engineering",
      "description": "Reducing the numerical format of weights or activations to decrease memory traffic and increase effective bandwidth.",
      "prerequisites": [
        "memory-wall"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "preemptive-scheduling",
      "name": "Preemptive Scheduling",
      "description": "Pausing long-running requests and swapping their state to CPU memory to allow shorter or higher-priority requests to execute.",
      "prerequisites": [
        "kv-cache"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "prefetch-buffer-depth",
      "name": "Prefetch Buffer Depth",
      "description": "The calculated number of prepared data batches required in host memory to fully absorb the worst-case I/O jitter.",
      "prerequisites": [
        "i-o-jitter",
        "host-dram-staging"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "prefill-decode-decomposition",
      "name": "Prefill-Decode Decomposition",
      "description": "The separation of LLM request processing into a compute-bound prompt processing phase and a memory-bound token generation phase.",
      "prerequisites": [
        "arithmetic-intensity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "prefix-caching",
      "name": "Prefix Caching",
      "description": "Sharing physical memory blocks for identical initial KV cache entries across different requests, such as common system prompts.",
      "prerequisites": [
        "pagedattention-mechanism"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "preprocessing-bottleneck",
      "name": "Preprocessing Bottleneck",
      "description": "A situation where CPU-bound data transformation dominates total request time, minimizing the benefits of GPU optimization.",
      "prerequisites": [
        "latency-budget"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "priority-flow-control",
      "name": "Priority Flow Control",
      "description": "A mechanism used in Ethernet to prevent buffer overflow by sending pause frames upstream to approximate lossless delivery.",
      "prerequisites": [
        "go-back-n",
        "roce"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "priority-inversion",
      "name": "Priority Inversion",
      "description": "A scenario where a high-priority job is indefinitely blocked because a low-priority job holding needed resources is stalled by medium-priority workloads.",
      "prerequisites": [
        "hold-and-wait-deadlock"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fleet_orchestration"
      ],
      "source_domains": [
        "fleet_orchestration"
      ],
      "question_count": 0
    },
    {
      "id": "process-group",
      "name": "Process Group",
      "description": "A specific subset of GPUs in a cluster that exclusively participates in a designated collective communication operation.",
      "prerequisites": [
        "3d-parallelism"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_collective_communication"
      ],
      "source_domains": [
        "collective_communication"
      ],
      "question_count": 0
    },
    {
      "id": "processing-elements",
      "name": "Processing Elements",
      "description": "The fundamental building blocks of AI accelerators that integrate dedicated computational units and local memory.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "source_domains": [
        "hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "progressive-deployment",
      "name": "Progressive Deployment",
      "description": "A hybrid architecture pattern where a massive cloud model is systematically compressed into smaller versions for edge and mobile tiers.",
      "prerequisites": [
        "hierarchical-processing",
        "model-quantization"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_systems"
      ],
      "source_domains": [
        "ml_systems"
      ],
      "question_count": 0
    },
    {
      "id": "quadratic-attention-bottleneck",
      "name": "Quadratic Attention Bottleneck",
      "description": "The severe computational and memory constraint during Transformer training where materializing the attention matrix scales operationally at O(N^2) with sequence length.",
      "prerequisites": [
        "dynamic-attention-routing"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [
        "nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "quality-of-service",
      "name": "Quality of Service",
      "description": "Mechanisms like Virtual Lanes or Traffic Classes that enforce fairness and bandwidth guarantees for different traffic types.",
      "prerequisites": [
        "bisection-bandwidth"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "quantization-aware-training",
      "name": "Quantization-Aware Training",
      "description": "Integrating simulated low-precision rounding noise directly into the training loop to allow the network to adapt to quantization errors.",
      "prerequisites": [
        "precision-engineering"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "rail-optimized-scheduling",
      "name": "Rail-Optimized Scheduling",
      "description": "Placing data-parallel replicas on specific node GPUs to ensure gradient synchronization traffic stays within independent network switch hierarchies.",
      "prerequisites": [
        "topology-aware-placement"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fleet_orchestration"
      ],
      "source_domains": [
        "fleet_orchestration"
      ],
      "question_count": 0
    },
    {
      "id": "rail-optimized-topology",
      "name": "Rail-Optimized Topology",
      "description": "A physical wiring pattern isolating tensor-parallel communication paths across nodes into dedicated switch networks.",
      "prerequisites": [
        "tensor-parallelism"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "random-iops",
      "name": "Random IOPS",
      "description": "The number of small, non-contiguous input/output operations per second a system can perform.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "reactive-scaling",
      "name": "Reactive Scaling",
      "description": "Provisioning fast-starting CPU instances dynamically in response to sudden spikes in user traffic.",
      "prerequisites": [
        "cost-per-inference"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "real-time-fairness-monitoring",
      "name": "Real-time Fairness Monitoring",
      "description": "Telemetry infrastructure computing rolling fairness metrics during live inference to detect demographic disparities.",
      "prerequisites": [
        "demographic-parity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_responsible_ai"
      ],
      "source_domains": [
        "responsible_ai"
      ],
      "question_count": 0
    },
    {
      "id": "remote-direct-memory-access",
      "name": "Remote Direct Memory Access",
      "description": "A networking technology that bypasses the operating system kernel, allowing network cards to read directly from memory to drastically reduce per-message latency.",
      "prerequisites": [],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_collective_communication",
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "collective_communication",
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "resource-fragmentation",
      "name": "Resource Fragmentation",
      "description": "Inefficiencies occurring when leftover node resources are too small for pending jobs but cannot be combined across nodes.",
      "prerequisites": [
        "bin-packing-heuristics"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fleet_orchestration"
      ],
      "source_domains": [
        "fleet_orchestration"
      ],
      "question_count": 0
    },
    {
      "id": "rest-protocol",
      "name": "REST Protocol",
      "description": "A text-based, stateless HTTP interface widely used for public APIs, which incurs high serialization overhead.",
      "prerequisites": [
        "serialization-overhead"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "reticle-limit",
      "name": "Reticle Limit",
      "description": "The maximum physical area of a single silicon die that photolithographic equipment can expose in a single manufacturing pass.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "reverse-linked-graph",
      "name": "Reverse-Linked Graph",
      "description": "A memory structure built during the forward pass linking tensors to their creating functions, maintaining the necessary state for backpropagation.",
      "prerequisites": [
        "reverse-mode-differentiation",
        "computational-graph"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "reverse-mode-differentiation",
      "name": "Reverse Mode Differentiation",
      "description": "An automatic differentiation technique that computes gradients backward from output to inputs, efficiently computing millions of parameter gradients in a single pass.",
      "prerequisites": [
        "automatic-differentiation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "ring-allreduce",
      "name": "Ring AllReduce",
      "description": "A bandwidth-optimized collective algorithm that arranges GPUs in a logical circle, breaking data into chunks to maximize network utilization without bottlenecking a single node.",
      "prerequisites": [
        "allreduce"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_collective_communication",
        "vol2_distributed_training"
      ],
      "source_domains": [
        "collective_communication",
        "distributed_training"
      ],
      "question_count": 0
    },
    {
      "id": "robust-optimization",
      "name": "Robust Optimization",
      "description": "Modifying the training objective with specialized loss functions to minimize the impact of outliers or poisoned data instances.",
      "prerequisites": [
        "adversarial-attacks"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_robust_ai"
      ],
      "source_domains": [
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "roce",
      "name": "RoCE",
      "description": "A protocol that implements RDMA semantics over standard Ethernet networks using UDP/IP encapsulation.",
      "prerequisites": [
        "remote-direct-memory-access"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "role-based-access-control",
      "name": "Role-Based Access Control",
      "description": "Security infrastructure mapping organizational policies into database permissions across distributed ML storage tiers.",
      "prerequisites": [
        "stationarity-assumption"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr"
      ],
      "source_domains": [
        "responsible_engr"
      ],
      "question_count": 0
    },
    {
      "id": "roofline-model",
      "name": "Roofline Model",
      "description": "An analytical framework that diagnoses whether a workload is compute-bound or memory-bound based on its arithmetic intensity and hardware limits.",
      "prerequisites": [
        "arithmetic-intensity",
        "memory-wall",
        "memory-boundedness"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration",
        "vol1_training",
        "vol2_compute_infrastructure",
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "compute_infrastructure",
        "hw_acceleration",
        "performance_engineering",
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "runtime-input-validation",
      "name": "Runtime Input Validation",
      "description": "Ensuring incoming data conforms to expected formats, statistical properties, or semantic constraints before reaching the model.",
      "prerequisites": [
        "layered-defense"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "runtime-output-monitoring",
      "name": "Runtime Output Monitoring",
      "description": "Detecting adversarial behavior or distribution shifts by analyzing a model's prediction confidence and response patterns in real time.",
      "prerequisites": [
        "layered-defense"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "saliency-maps",
      "name": "Saliency Maps",
      "description": "Lightweight explanation methods using single backward passes on input gradients, suitable for latency-sensitive edge environments.",
      "prerequisites": [
        "system-explainability"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_responsible_ai"
      ],
      "source_domains": [
        "responsible_ai"
      ],
      "question_count": 0
    },
    {
      "id": "scaling-efficiency-curve",
      "name": "Scaling Efficiency Curve",
      "description": "An empirical measurement of how a job's throughput scales with worker count, used by schedulers to prevent allocating GPUs that yield diminishing returns.",
      "prerequisites": [
        "static-batching",
        "elastic-training"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fleet_orchestration",
        "vol2_inference"
      ],
      "source_domains": [
        "fleet_orchestration",
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "schema-validation",
      "name": "Schema Validation",
      "description": "The synchronous enforcement of expected data types and value ranges to reject malformed records entering the pipeline.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering"
      ],
      "source_domains": [
        "data_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "secure-aggregation",
      "name": "Secure Aggregation",
      "description": "Cryptographic protocols that allow a server to sum distributed model updates without revealing any individual edge device's specific parameter contributions.",
      "prerequisites": [
        "federated-learning",
        "gradient-inversion-attack"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_edge_intelligence",
        "vol2_security_privacy"
      ],
      "source_domains": [
        "edge_intelligence",
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "secure-model-deployment",
      "name": "Secure Model Deployment",
      "description": "Applying packaging, encryption, and access controls to protect a trained model during inference operations.",
      "prerequisites": [
        "layered-defense"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "secure-model-design",
      "name": "Secure Model Design",
      "description": "Early-stage architectural choices, such as reducing capacity or adding abstention, to reduce a model's exposure to attacks.",
      "prerequisites": [
        "layered-defense"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "secure-multiparty-computation",
      "name": "Secure Multiparty Computation",
      "description": "Distributing computation across multiple parties such that no single entity learns the complete input or output data.",
      "prerequisites": [
        "confidence-calibration"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "self-supervised-learning",
      "name": "Self-Supervised Learning",
      "description": "A learning paradigm that discovers representations by solving pretext tasks, building inherent resilience against distribution shifts.",
      "prerequisites": [
        "on-device-learning"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_edge_intelligence",
        "vol2_robust_ai"
      ],
      "source_domains": [
        "edge_intelligence",
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "sequential-dependency",
      "name": "Sequential Dependency",
      "description": "The inherent hardware constraint in recurrent networks where time step computations cannot be parallelized because each step waits on the previous step's hidden state.",
      "prerequisites": [
        "dense-matrix-multiplication"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [
        "nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "sequential-streaming-formats",
      "name": "Sequential Streaming Formats",
      "description": "Bespoke data serialization formats that eliminate internal indices and per-sample metadata to maximize sequential read bandwidth.",
      "prerequisites": [
        "metadata-overhead",
        "data-sharding"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "sequential-throughput",
      "name": "Sequential Throughput",
      "description": "The rate at which data is transferred when read contiguously from storage, maximizing underlying hardware efficiency.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "serialization-overhead",
      "name": "Serialization Overhead",
      "description": "The computational latency incurred by converting data formats, such as parsing JSON text into machine-native types.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "serializer-deserializer",
      "name": "Serializer/Deserializer",
      "description": "A circuit that converts parallel data from a switch ASIC into a high-speed serial stream for transmission.",
      "prerequisites": [
        "bisection-bandwidth"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "service-level-objective",
      "name": "Service Level Objective",
      "description": "A formal target for acceptable latency that dictates architectural decisions in a serving system.",
      "prerequisites": [
        "latency",
        "tail-latency"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "serving-autoscaling",
      "name": "Serving Autoscaling",
      "description": "Dynamically scaling horizontal inference capacity to handle fluctuating traffic demands without violating latency objectives.",
      "prerequisites": [
        "sutva-violations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_ops_scale"
      ],
      "source_domains": [
        "ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "serving-cost-dominance-law",
      "name": "Serving Cost Dominance Law",
      "description": "The economic principle that the ongoing operational costs of inference exceed the initial capital expenditure of training by orders of magnitude.",
      "prerequisites": [
        "kv-cache"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "serving-inversion",
      "name": "Serving Inversion",
      "description": "The shift in priority from maximizing throughput during training to minimizing latency during production serving.",
      "prerequisites": [
        "throughput",
        "latency"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "serving-tax",
      "name": "Serving Tax",
      "description": "The cumulative latency overhead introduced by distributing inference, including network communication, serialization, and coordination.",
      "prerequisites": [
        "kv-cache"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "shadow-deployment",
      "name": "Shadow Deployment",
      "description": "Executing a candidate model on real-world traffic in parallel with production without exposing its predictions to end users.",
      "prerequisites": [
        "hardware-memory-hierarchy"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_ops",
        "vol2_ops_scale"
      ],
      "source_domains": [
        "ml_ops",
        "ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "shadow-model-evaluation",
      "name": "Shadow Model Evaluation",
      "description": "Simultaneously running a frozen baseline model alongside a newly adapted variant to detect performance degradation directly on real-world data streams.",
      "prerequisites": [
        "on-device-learning"
      ],
      "tracks": [
        "edge"
      ],
      "source_chapters": [
        "vol2_edge_intelligence"
      ],
      "source_domains": [
        "edge_intelligence"
      ],
      "question_count": 0
    },
    {
      "id": "shard-level-shuffling",
      "name": "Shard-Level Shuffling",
      "description": "An algorithm that randomizes the order of large data blocks and shuffles samples within them to approximate randomness while preserving sequential I/O.",
      "prerequisites": [
        "data-sharding"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "sharded-checkpointing",
      "name": "Sharded Checkpointing",
      "description": "A distributed checkpointing method where each worker writes only its assigned parameter and optimizer shard.",
      "prerequisites": [
        "distributed-checkpointing"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "side-channel-attack",
      "name": "Side-Channel Attack",
      "description": "Exploiting information inadvertently revealed through physical system implementation, such as power consumption or electromagnetic emissions.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "silent-data-corruption",
      "name": "Silent Data Corruption",
      "description": "A type of Byzantine failure where memory or computation errors produce wrong values without triggering explicit hardware or software exceptions.",
      "prerequisites": [
        "transient-failures",
        "byzantine-failures"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance",
        "vol2_robust_ai"
      ],
      "source_domains": [
        "fault_tolerance",
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "silent-degradation",
      "name": "Silent Degradation",
      "description": "A production failure mode where deployed models produce plausible but incorrect outputs without generating explicit error signals.",
      "prerequisites": [
        "system-benchmarking"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_benchmarking"
      ],
      "source_domains": [
        "benchmarking"
      ],
      "question_count": 0
    },
    {
      "id": "silent-model-failures",
      "name": "Silent Model Failures",
      "description": "Gradual degradation of predictive accuracy due to real-world changes that go undetected by traditional infrastructure health metrics.",
      "prerequisites": [
        "distribution-shift"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_ops"
      ],
      "source_domains": [
        "ml_ops"
      ],
      "question_count": 0
    },
    {
      "id": "silicon-interposer",
      "name": "Silicon Interposer",
      "description": "A passive silicon substrate with microscopic etched wiring layers that enables thousands of parallel connections between the processor die and memory stacks.",
      "prerequisites": [
        "high-bandwidth-memory"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "simd-execution",
      "name": "SIMD Execution",
      "description": "A hardware execution model that applies identical operations to multiple data elements in parallel to maximize throughput.",
      "prerequisites": [
        "vector-operations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "source_domains": [
        "hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "simt-execution-model",
      "name": "SIMT Execution Model",
      "description": "A parallel computing architecture where a single instruction is mapped across a massive grid of threads to hide memory latency.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "single-root-i-o-virtualization",
      "name": "Single Root I/O Virtualization",
      "description": "A hardware virtualization standard allowing a single physical NIC to present itself as multiple independent virtual functions.",
      "prerequisites": [
        "remote-direct-memory-access"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "single-static-assignment",
      "name": "Single Static Assignment",
      "description": "An intermediate representation format where each variable is assigned exactly once, simplifying dependency analysis for framework compilers.",
      "prerequisites": [
        "just-in-time-compilation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "skip-connections",
      "name": "Skip Connections",
      "description": "Architectural bypass pathways that sum outputs across multiple layers to allow gradients to flow unimpeded, directly solving the vanishing gradient problem in deep networks.",
      "prerequisites": [
        "dense-matrix-multiplication"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [
        "nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "small-file-problem",
      "name": "Small File Problem",
      "description": "A severe performance collapse in parallel file systems caused by reading millions of tiny files that overwhelm metadata servers.",
      "prerequisites": [
        "random-iops",
        "metadata-overhead"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage"
      ],
      "question_count": 0
    },
    {
      "id": "smoothquant",
      "name": "SmoothQuant",
      "description": "A technique that migrates quantization difficulty from dynamic activations to static weights via a mathematically equivalent per-channel scaling transformation.",
      "prerequisites": [
        "precision-engineering",
        "block-wise-quantization"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "sparse-embedding-lookups",
      "name": "Sparse Embedding Lookups",
      "description": "Index-based memory operations in recommendation models that retrieve dense vectors for categorical features, creating random memory access patterns with zero arithmetic intensity.",
      "prerequisites": [
        "memory-capacity-bound-workload"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [
        "nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "sparse-gradient-updates",
      "name": "Sparse Gradient Updates",
      "description": "Exploiting natural sparsity by transmitting only the largest or most significant gradient components to further minimize communication overhead.",
      "prerequisites": [
        "federated-learning"
      ],
      "tracks": [
        "edge"
      ],
      "source_chapters": [
        "vol2_edge_intelligence"
      ],
      "source_domains": [
        "edge_intelligence"
      ],
      "question_count": 0
    },
    {
      "id": "sparse-scatter",
      "name": "Sparse Scatter",
      "description": "A workload archetype with irregular memory access patterns and poor cache locality, constrained by memory capacity and access latency.",
      "prerequisites": [
        "memory-bound-workload"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_systems"
      ],
      "source_domains": [
        "ml_systems"
      ],
      "question_count": 0
    },
    {
      "id": "spatial-and-temporal-locality",
      "name": "Spatial and Temporal Locality",
      "description": "An operating system assumption of predictable memory access patterns that stochastic ML data loading workloads fundamentally violate.",
      "prerequisites": [
        "data-cascade"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_workflow"
      ],
      "source_domains": [
        "ml_workflow"
      ],
      "question_count": 0
    },
    {
      "id": "special-function-units",
      "name": "Special Function Units",
      "description": "Dedicated hardware circuits designed to execute non-linear mathematical operations like activations and normalizations efficiently.",
      "prerequisites": [
        "matrix-operations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "source_domains": [
        "hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "spectral-signatures",
      "name": "Spectral Signatures",
      "description": "A statistical defense technique that detects poisoned training data by analyzing the covariance matrix of learned representations for anomalies.",
      "prerequisites": [
        "data-poisoning"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_robust_ai"
      ],
      "source_domains": [
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "speculative-decoding",
      "name": "Speculative Decoding",
      "description": "An algorithm where a small draft model generates candidate tokens sequentially and a large target model verifies them in parallel.",
      "prerequisites": [
        "cuda-streams"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_inference",
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "inference",
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "speculative-execution-flaws",
      "name": "Speculative Execution Flaws",
      "description": "Hardware design vulnerabilities that exploit processor performance optimizations to leak sensitive data through microarchitectural side channels.",
      "prerequisites": [
        "confidence-calibration"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "spot-instances",
      "name": "Spot Instances",
      "description": "Excess cloud compute capacity offered at steep discounts that the provider can preemptively reclaim with minimal warning.",
      "prerequisites": [
        "fleet-orchestration"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fleet_orchestration"
      ],
      "source_domains": [
        "fleet_orchestration"
      ],
      "question_count": 0
    },
    {
      "id": "spurious-correlations",
      "name": "Spurious Correlations",
      "description": "Features that are statistically predictive in the training dataset but lack a causal relationship with the target label.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_robust_ai"
      ],
      "source_domains": [
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "sram",
      "name": "SRAM",
      "description": "Static RAM — fast, expensive on-chip memory used for caches and tensor arenas.",
      "prerequisites": [
        "matrix-operations"
      ],
      "tracks": [
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "stage-interface-contracts",
      "name": "Stage Interface Contracts",
      "description": "Explicit input requirements, output deliverables, and quality invariants that must hold before advancing to the next lifecycle stage.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_workflow"
      ],
      "source_domains": [
        "ml_workflow"
      ],
      "question_count": 0
    },
    {
      "id": "staged-system-pipeline",
      "name": "Staged System Pipeline",
      "description": "A coordinated architecture of data loading, host-to-device transfer, computation, and parameter synchronization.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "staleness-cost-function",
      "name": "Staleness Cost Function",
      "description": "An economic model that quantifies the financial impact of model accuracy decay over time to determine optimal retraining intervals.",
      "prerequisites": [
        "continuous-training-pipelines"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_ops"
      ],
      "source_domains": [
        "ml_ops"
      ],
      "question_count": 0
    },
    {
      "id": "state-reconciliation",
      "name": "State Reconciliation",
      "description": "Synchronization protocols enabling clients returning from extended offline periods to safely integrate their local adaptations with new global model states.",
      "prerequisites": [
        "federated-learning"
      ],
      "tracks": [
        "edge"
      ],
      "source_chapters": [
        "vol2_edge_intelligence"
      ],
      "source_domains": [
        "edge_intelligence"
      ],
      "question_count": 0
    },
    {
      "id": "stateful-serving",
      "name": "Stateful Serving",
      "description": "An inference architecture that maintains context across a sequence of requests, complicating failover and recovery.",
      "prerequisites": [
        "fault-tolerance"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance",
        "vol2_inference"
      ],
      "source_domains": [
        "fault_tolerance",
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "stateless-serving",
      "name": "Stateless Serving",
      "description": "An inference architecture where each request is independent and no per-session context is maintained across requests.",
      "prerequisites": [
        "fault-tolerance"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "static-batching",
      "name": "Static Batching",
      "description": "Accumulating a fixed number of inference requests before dispatching them to the accelerator to maximize throughput.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "static-computation-graphs",
      "name": "Static Computation Graphs",
      "description": "A define-then-run execution model where the entire network architecture is symbolically defined and optimized before any actual computation occurs.",
      "prerequisites": [
        "computational-graph"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "static-inference",
      "name": "Static Inference",
      "description": "Pre-computing predictions for anticipated inputs and storing them in a lookup table to eliminate serving latency.",
      "prerequisites": [
        "inference-pipeline"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "static-power",
      "name": "Static Power",
      "description": "Energy consumed through leakage current that flows even when transistors are inactive and increases exponentially with temperature.",
      "prerequisites": [
        "cmos-power-equation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "stationarity-assumption",
      "name": "Stationarity Assumption",
      "description": "The foundational assumption in supervised learning that training and deployment data distributions must match.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr"
      ],
      "source_domains": [
        "responsible_engr"
      ],
      "question_count": 0
    },
    {
      "id": "statistical-parity-difference",
      "name": "Statistical Parity Difference",
      "description": "The absolute difference in positive outcome selection rates between privileged and unprivileged groups.",
      "prerequisites": [
        "demographic-parity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_responsible_ai"
      ],
      "source_domains": [
        "responsible_ai"
      ],
      "question_count": 0
    },
    {
      "id": "sticky-routing",
      "name": "Sticky Routing",
      "description": "Using consistent hashing to guarantee a user's requests route to the exact same model version during distributed deployment transitions.",
      "prerequisites": [
        "multi-region-deployment"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_ops_scale"
      ],
      "source_domains": [
        "ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "store-and-forward-architecture",
      "name": "Store-and-forward Architecture",
      "description": "A network transmission strategy for intermittent connectivity that batches data locally and transmits it asynchronously during connectivity windows.",
      "prerequisites": [
        "data-cascade"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_workflow"
      ],
      "source_domains": [
        "ml_workflow"
      ],
      "question_count": 0
    },
    {
      "id": "straggler-mitigation",
      "name": "Straggler Mitigation",
      "description": "Asynchronous device coordination strategies that ensure slow or poorly connected edge devices do not stall the global training progress.",
      "prerequisites": [
        "fault-tolerance",
        "federated-learning"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_edge_intelligence",
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "edge_intelligence",
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "stream-ingestion",
      "name": "Stream Ingestion",
      "description": "Processing event data continuously in real-time to minimize latency, requiring always-on infrastructure and complex backpressure handling.",
      "prerequisites": [
        "batch-ingestion"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering"
      ],
      "source_domains": [
        "data_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "streaming-inference",
      "name": "Streaming Inference",
      "description": "Processing continuous input data incrementally as it arrives with strict real-time latency constraints instead of batching.",
      "prerequisites": [
        "dynamic-batching"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "structured-sparsity",
      "name": "Structured Sparsity",
      "description": "A predictable pattern of zero values that allows hardware accelerators to efficiently compress and compute sparse matrices without irregular memory access penalties.",
      "prerequisites": [
        "matrix-operations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "source_domains": [
        "hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "sutva-violations",
      "name": "SUTVA Violations",
      "description": "The failure of A/B testing independence assumptions caused when networked users influence control group outcomes.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_ops_scale"
      ],
      "source_domains": [
        "ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "synchronous-checkpointing",
      "name": "Synchronous Checkpointing",
      "description": "A mechanism where all workers coordinate to reach a consistent state and pause training while data is written to storage.",
      "prerequisites": [
        "checkpointing"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "synthetic-data-generation",
      "name": "Synthetic Data Generation",
      "description": "Training generative models to produce artificial data that preserves statistical properties of real data without retaining personally identifiable information.",
      "prerequisites": [
        "confidence-calibration"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "source_domains": [
        "security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "system-benchmarking",
      "name": "System Benchmarking",
      "description": "The evaluation of whether hardware delivers promised performance under realistic ML workloads.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_benchmarking"
      ],
      "source_domains": [
        "benchmarking"
      ],
      "question_count": 0
    },
    {
      "id": "system-entropy",
      "name": "System Entropy",
      "description": "The continuous statistical degradation of deployed model accuracy as real-world data drifts away from the training distribution.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_systems"
      ],
      "source_domains": [
        "ml_systems"
      ],
      "question_count": 0
    },
    {
      "id": "system-explainability",
      "name": "System Explainability",
      "description": "The mechanism by which the control plane exposes internal model state to human operators for debugging and verification.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_responsible_ai"
      ],
      "source_domains": [
        "responsible_ai"
      ],
      "question_count": 0
    },
    {
      "id": "systolic-array",
      "name": "Systolic Array",
      "description": "A fixed grid of compute units where data flows between neighboring cells in a pipelined, wave-like pattern to eliminate instruction overhead.",
      "prerequisites": [
        "matrix-operations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration",
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure",
        "hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "tail-latency",
      "name": "Tail Latency",
      "description": "The response time experienced by a small percentage of outliers, such as the 99th percentile, which dictates the median user experience at scale.",
      "prerequisites": [
        "latency"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving",
        "vol2_data_storage"
      ],
      "source_domains": [
        "data_storage",
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "telemetry-aggregation",
      "name": "Telemetry Aggregation",
      "description": "Categorizing and sampling logging data at the edge to prevent monitoring network bottlenecks at high throughput.",
      "prerequisites": [
        "sutva-violations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_ops_scale"
      ],
      "source_domains": [
        "ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "tensor-abstraction",
      "name": "Tensor Abstraction",
      "description": "The universal multi-dimensional data container that encodes mathematical semantics, hardware device placement, and memory layout for framework operations.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "tensor-arena",
      "name": "Tensor Arena",
      "description": "A fixed, pre-allocated memory region used in bare-metal serving where dynamic memory allocation is impossible.",
      "prerequisites": [
        "bare-metal-serving"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "tensor-cores",
      "name": "Tensor Cores",
      "description": "Specialized hardware units optimized for executing multi-dimensional matrix multiplications and accumulations on fixed-size blocks.",
      "prerequisites": [
        "matrix-operations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration",
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure",
        "hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "tensor-format-transformation",
      "name": "Tensor Format Transformation",
      "description": "Converting network payloads into optimized memory layouts required by the model runtime.",
      "prerequisites": [
        "inference-server"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "tensor-parallelism",
      "name": "Tensor Parallelism",
      "description": "Splitting individual matrix multiplication operations across multiple accelerators to reduce per-device memory load and compute time.",
      "prerequisites": [
        "model-parallelism",
        "model-sharding"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training",
        "vol2_distributed_training",
        "vol2_inference",
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "distributed_training",
        "inference",
        "network_fabrics",
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "tensor-stride-patterns",
      "name": "Tensor Stride Patterns",
      "description": "The mathematical mapping between multi-dimensional tensor indices and linear physical memory addresses, critical for optimizing cache utilization and data layout.",
      "prerequisites": [
        "tensor-abstraction"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "source_domains": [
        "frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "tensorrt",
      "name": "TensorRT",
      "description": "NVIDIA inference optimizer that fuses layers, quantizes, and generates optimized kernels.",
      "prerequisites": [
        "operator-fusion",
        "model-quantization"
      ],
      "tracks": [
        "cloud",
        "edge"
      ],
      "source_chapters": [
        "vol1_model_serving",
        "vol2_inference"
      ],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "test-time-compute-scaling",
      "name": "Test-Time Compute Scaling",
      "description": "Dynamically scaling the amount of computation spent per inference request proportional to the complexity of the task.",
      "prerequisites": [
        "kv-cache"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "tflite",
      "name": "TFLite",
      "description": "TensorFlow Lite — mobile/edge inference framework with delegate system.",
      "prerequisites": [
        "on-device-learning"
      ],
      "tracks": [
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_edge_intelligence"
      ],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "thermal-design-power",
      "name": "Thermal Design Power",
      "description": "The maximum sustained heat dissipation generated by a processor that the physical infrastructure must be designed to continuously cool.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "threshold-adjustment",
      "name": "Threshold Adjustment",
      "description": "A post-hoc fairness mitigation technique that applies different decision boundaries to different demographic groups to equalize outcomes.",
      "prerequisites": [
        "equality-of-opportunity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr"
      ],
      "source_domains": [
        "responsible_engr"
      ],
      "question_count": 0
    },
    {
      "id": "through-silicon-vias",
      "name": "Through-Silicon Vias",
      "description": "Microscopic vertical copper connections etched directly through stacked memory dies to drastically shorten signal paths and increase bandwidth.",
      "prerequisites": [
        "high-bandwidth-memory"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "throughput",
      "name": "Throughput",
      "description": "The rate at which a system processes requests, measured in samples per second.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "tiered-storage",
      "name": "Tiered Storage",
      "description": "A data architecture that dynamically migrates datasets across hot, warm, and cold storage tiers based on access recency.",
      "prerequisites": [
        "data-cascade"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_workflow"
      ],
      "source_domains": [
        "ml_workflow"
      ],
      "question_count": 0
    },
    {
      "id": "tiling-algorithms",
      "name": "Tiling Algorithms",
      "description": "Decomposing large matrix computations into smaller blocks that fit entirely within fast on-chip SRAM to minimize HBM round-trips.",
      "prerequisites": [
        "io-aware-algorithm-design"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "time-per-output-token",
      "name": "Time Per Output Token",
      "description": "The latency required to generate each subsequent token during the memory-bandwidth-bound decode phase.",
      "prerequisites": [
        "time-to-first-token"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "time-to-first-token",
      "name": "Time to First Token",
      "description": "The latency required to process an input prompt and generate the initial output token during the compute-bound prefill phase.",
      "prerequisites": [
        "model-benchmarking"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_benchmarking",
        "vol2_inference"
      ],
      "source_domains": [
        "benchmarking",
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "timeline-profiling",
      "name": "Timeline Profiling",
      "description": "Visualizing the chronological sequence of CPU and GPU execution events to identify idle gaps and bottlenecks.",
      "prerequisites": [
        "application-embedded-serving"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "tinyml",
      "name": "TinyML",
      "description": "A deployment paradigm for microcontrollers operating under extreme milliwatt power budgets and kilobyte memory limits.",
      "prerequisites": [
        "memory-wall"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_systems"
      ],
      "source_domains": [
        "ml_systems"
      ],
      "question_count": 0
    },
    {
      "id": "top-of-rack-switch",
      "name": "Top-of-Rack Switch",
      "description": "The fundamental physical aggregation point connecting servers within a rack, defining the minimum failure domain.",
      "prerequisites": [
        "bisection-bandwidth"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "topology-aware-placement",
      "name": "Topology-Aware Placement",
      "description": "Assigning distributed workloads to nodes based on physical network hierarchy to confine high-bandwidth communication within local switches.",
      "prerequisites": [
        "fleet-orchestration"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fleet_orchestration"
      ],
      "source_domains": [
        "fleet_orchestration"
      ],
      "question_count": 0
    },
    {
      "id": "torchdynamo",
      "name": "TorchDynamo",
      "description": "A Python bytecode-level graph capture tool that traces model execution while gracefully handling dynamic Python control flow.",
      "prerequisites": [
        "graph-compilation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "source_domains": [
        "performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "torus-topology",
      "name": "Torus Topology",
      "description": "A grid-based topology where each node connects directly to its geometric neighbors, offering high local but limited global bandwidth.",
      "prerequisites": [
        "bisection-bandwidth"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "total-cost-of-ownership",
      "name": "Total Cost of Ownership",
      "description": "The comprehensive financial cost of an ML system, typically heavily dominated by inference compute over its operational lifetime.",
      "prerequisites": [
        "stationarity-assumption"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr"
      ],
      "source_domains": [
        "responsible_engr"
      ],
      "question_count": 0
    },
    {
      "id": "train-serve-split",
      "name": "Train-Serve Split",
      "description": "A hybrid architecture pattern where computationally intensive training occurs in the cloud while inference runs locally.",
      "prerequisites": [
        "data-gravity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_systems"
      ],
      "source_domains": [
        "ml_systems"
      ],
      "question_count": 0
    },
    {
      "id": "training-cost-asymmetry",
      "name": "Training Cost Asymmetry",
      "description": "The million-to-one resource disparity between training a model and running inference due to repeated forward passes, backward passes, and optimizer steps.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "training-serving-skew",
      "name": "Training-Serving Skew",
      "description": "The silent degradation of model accuracy caused by applying different feature computation logic during training versus production.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering",
        "vol1_ml_ops"
      ],
      "source_domains": [
        "data_engineering",
        "ml_ops"
      ],
      "question_count": 0
    },
    {
      "id": "transformation-lineage",
      "name": "Transformation Lineage",
      "description": "Systematically tracking the sequence, parameters, and versions of code applied to raw data to guarantee reproducibility.",
      "prerequisites": [
        "idempotent-transformations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering"
      ],
      "source_domains": [
        "data_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "transformer-architecture",
      "name": "Transformer Architecture",
      "description": "Self-attention-based architecture dominating NLP and increasingly vision.",
      "prerequisites": [
        "dense-matrix-multiplication"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "transient-failures",
      "name": "Transient Failures",
      "description": "Temporary hardware disruptions like single-event upsets or voltage fluctuations that cause incorrect computations without permanent hardware damage.",
      "prerequisites": [
        "fault-tolerance"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "transient-loads",
      "name": "Transient Loads",
      "description": "Fluctuating power demands caused by synchronized cluster operations that require energy buffering to maintain grid stability.",
      "prerequisites": [
        "cmos-power-equation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "source_domains": [
        "sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "translation-equivariance",
      "name": "Translation Equivariance",
      "description": "An architectural property achieved via parameter sharing where shifting the input produces a structurally corresponding shift in the output feature map.",
      "prerequisites": [
        "dense-matrix-multiplication"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [
        "nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "ultra-ethernet",
      "name": "Ultra Ethernet",
      "description": "An emerging Ethernet standard designed for AI workloads featuring native packet spraying and in-order delivery.",
      "prerequisites": [
        "roce"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "source_domains": [
        "network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "vanishing-gradient-problem",
      "name": "Vanishing Gradient Problem",
      "description": "A mathematical failure mode where error signals shrink exponentially through deep layers due to saturated activation functions, silently preventing early layers from updating.",
      "prerequisites": [
        "activation-memory-footprint"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_computation"
      ],
      "source_domains": [
        "nn_computation"
      ],
      "question_count": 0
    },
    {
      "id": "vector-operations",
      "name": "Vector Operations",
      "description": "Computations that apply a single mathematical instruction to multiple independent data elements simultaneously.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "source_domains": [
        "hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "wafer-scale-engine",
      "name": "Wafer-Scale Engine",
      "description": "A massive, continuous piece of silicon containing hundreds of thousands of cores connected by an ultra-fast on-chip fabric to minimize communication latency.",
      "prerequisites": [
        "custom-asic"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "warm-restart",
      "name": "Warm Restart",
      "description": "A recovery procedure that preserves the in-memory state of surviving workers and only replaces the failed nodes.",
      "prerequisites": [
        "checkpointing"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "warp-divergence",
      "name": "Warp Divergence",
      "description": "The performance degradation that occurs when threads within the same execution group take different branching paths and must run sequentially.",
      "prerequisites": [
        "simt-execution-model"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "weight-stationary",
      "name": "Weight Stationary",
      "description": "A dataflow strategy that keeps model parameters fixed in local memory while streaming inputs and accumulating partial sums.",
      "prerequisites": [
        "dataflow-optimization"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "source_domains": [
        "hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "young-daly-formula",
      "name": "Young-Daly Formula",
      "description": "A mathematical equation used to calculate the optimal checkpointing frequency by balancing save overhead against potential rework time.",
      "prerequisites": [
        "failure-rate",
        "checkpointing"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_distributed_training",
        "vol2_fault_tolerance"
      ],
      "source_domains": [
        "distributed_training",
        "fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "zero-copy-inference",
      "name": "Zero-Copy Inference",
      "description": "Eliminating CPU memory copies by allowing accelerators to read data directly from hardware buffers like camera sensors.",
      "prerequisites": [
        "application-embedded-serving"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "zero-copy-loading",
      "name": "Zero-Copy Loading",
      "description": "Using memory-mapped files to instantly access model weights directly from disk without unpickling or copying to RAM.",
      "prerequisites": [
        "application-embedded-serving"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "zero-copy-serialization",
      "name": "Zero-Copy Serialization",
      "description": "Using binary formats like FlatBuffers or Protobuf that map directly to memory structures, avoiding deserialization processing.",
      "prerequisites": [
        "serialization-overhead"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "source_domains": [
        "model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "zero-optimizations",
      "name": "ZeRO Optimizations",
      "description": "Memory efficiency techniques that partition model states to prevent data parallelism from failing on large parameter counts.",
      "prerequisites": [
        "3d-parallelism"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_distributed_training"
      ],
      "source_domains": [
        "distributed_training"
      ],
      "question_count": 0
    },
    {
      "id": "low-rank-adaptation",
      "name": "Low-Rank Adaptation",
      "description": "Parameter-efficient fine-tuning by decomposing weight updates into low-rank matrices, reducing trainable parameters by orders of magnitude.",
      "prerequisites": [
        "gradient-accumulation",
        "mixed-precision-training"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "fully-sharded-data-parallel",
      "name": "Fully Sharded Data Parallel",
      "description": "A memory-efficient distributed training strategy that shards model parameters, gradients, and optimizer states across workers, reconstructing them on-demand during forward and backward passes.",
      "prerequisites": [
        "allreduce",
        "model-sharding",
        "data-parallelism"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_distributed_training"
      ],
      "source_domains": [
        "distributed_training"
      ],
      "question_count": 0
    },
    {
      "id": "flash-decoding",
      "name": "Flash Decoding",
      "description": "A parallelization strategy for autoregressive decoding that splits the KV cache across threads to maximize GPU utilization during the memory-bound decode phase.",
      "prerequisites": [
        "memory-bound-workload",
        "autoregressive-kv-cache",
        "flash-attention"
      ],
      "tracks": [
        "cloud",
        "edge"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "chunked-prefill",
      "name": "Chunked Prefill",
      "description": "Splitting the compute-heavy prefill phase of LLM inference into smaller chunks that can be interleaved with decode requests, preventing head-of-line blocking.",
      "prerequisites": [
        "continuous-batching",
        "preemptive-scheduling"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "multi-query-attention",
      "name": "Multi-Query Attention",
      "description": "An attention variant that shares key and value projections across all attention heads, dramatically reducing KV cache memory and decode-phase memory bandwidth.",
      "prerequisites": [
        "autoregressive-kv-cache",
        "grouped-query-attention"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [
        "nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "sliding-window-attention",
      "name": "Sliding Window Attention",
      "description": "A local attention pattern where each token attends only to a fixed-size window of preceding tokens, reducing the quadratic cost of full attention to linear.",
      "prerequisites": [
        "quadratic-attention-bottleneck",
        "flash-attention"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [
        "nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "sparse-attention",
      "name": "Sparse Attention",
      "description": "Attention mechanisms that compute only a subset of the full attention matrix by exploiting structural or learned sparsity patterns, trading approximation for reduced compute and memory.",
      "prerequisites": [
        "quadratic-attention-bottleneck",
        "activation-sparsity"
      ],
      "tracks": [
        "cloud",
        "edge"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [
        "nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "gradient-clipping",
      "name": "Gradient Clipping",
      "description": "Bounding gradient norms during backpropagation to prevent exploding gradients and stabilize training dynamics, especially critical for deep networks and mixed-precision training.",
      "prerequisites": [
        "mixed-precision-training",
        "vanishing-gradient-problem"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "dynamic-loss-scaling",
      "name": "Dynamic Loss Scaling",
      "description": "Automatically scaling the loss value before backpropagation in mixed-precision training to prevent gradient underflow in FP16, then unscaling before the optimizer step.",
      "prerequisites": [
        "mixed-precision-training",
        "gradient-clipping"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "learning-rate-scheduling",
      "name": "Learning Rate Scheduling",
      "description": "Systematic adjustment of the learning rate during training using strategies like warmup, cosine decay, or step decay to balance convergence speed and final model quality.",
      "prerequisites": [
        "gradient-accumulation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "adapter-layers",
      "name": "Adapter Layers",
      "description": "Small trainable bottleneck modules inserted between frozen pretrained layers, enabling parameter-efficient transfer learning by training less than 1% of total parameters.",
      "prerequisites": [
        "low-rank-adaptation",
        "knowledge-distillation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "gpu-virtualization",
      "name": "GPU Virtualization",
      "description": "Hardware and software techniques (MIG, MPS, vGPU) for partitioning a single physical GPU into isolated instances, enabling multi-tenant inference and improving utilization.",
      "prerequisites": [
        "capacity-planning",
        "bin-packing-heuristics"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "cxl-memory-expansion",
      "name": "CXL Memory Expansion",
      "description": "Compute Express Link protocol enabling cache-coherent memory pooling and expansion beyond local DRAM, offering a new tier in the memory hierarchy between DDR and network-attached storage.",
      "prerequisites": [
        "cache-hierarchy",
        "tiered-storage",
        "memory-wall"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "tokenization",
      "name": "Tokenization",
      "description": "The process of converting raw text into discrete token IDs using algorithms like BPE or SentencePiece, which directly determines vocabulary size, sequence length, and thus compute and memory requirements.",
      "prerequisites": [
        "transformer-architecture"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "source_domains": [
        "nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "red-teaming",
      "name": "Red Teaming",
      "description": "Systematic adversarial testing of ML models by human evaluators or automated methods to discover failure modes, safety violations, and unexpected behaviors before deployment.",
      "prerequisites": [
        "adversarial-inputs",
        "certified-robustness"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol2_robust_ai"
      ],
      "source_domains": [
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "inference-guardrails",
      "name": "Inference Guardrails",
      "description": "Input and output filtering systems that enforce safety, compliance, and quality constraints on model responses at serving time, adding latency but preventing harmful outputs.",
      "prerequisites": [
        "service-level-objective",
        "red-teaming"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol2_robust_ai"
      ],
      "source_domains": [
        "robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "request-routing",
      "name": "Request Routing",
      "description": "Intelligent dispatching of inference requests across heterogeneous serving backends based on model version, request complexity, latency budget, or cost constraints.",
      "prerequisites": [
        "capacity-planning",
        "preemptive-scheduling",
        "load-balancer"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "token-budget",
      "name": "Token Budget",
      "description": "The maximum number of input plus output tokens allowed per request, which directly bounds memory (KV cache), compute (attention FLOPs), and cost, serving as the primary resource constraint in LLM serving.",
      "prerequisites": [
        "cost-per-inference",
        "autoregressive-kv-cache"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol2_inference"
      ],
      "source_domains": [
        "inference"
      ],
      "question_count": 0
    },
    {
      "id": "deepspeed-zero",
      "name": "DeepSpeed ZeRO",
      "description": "A family of memory optimization stages (ZeRO-1/2/3) that progressively shard optimizer states, gradients, and parameters across data-parallel workers, trading communication volume for per-device memory reduction.",
      "prerequisites": [
        "gradient-accumulation",
        "fully-sharded-data-parallel",
        "data-parallelism"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_distributed_training"
      ],
      "source_domains": [
        "distributed_training"
      ],
      "question_count": 0
    },
    {
      "id": "megatron-parallelism",
      "name": "Megatron-Style Tensor Parallelism",
      "description": "A tensor parallelism strategy that partitions weight matrices along column and row dimensions across devices, with conjugate all-reduce operations to reconstruct activations, enabling linear scaling of model width.",
      "prerequisites": [
        "tensor-parallelism",
        "allreduce",
        "3d-parallelism"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_distributed_training"
      ],
      "source_domains": [
        "distributed_training"
      ],
      "question_count": 0
    },
    {
      "id": "gptq-quantization",
      "name": "GPTQ Quantization",
      "description": "A post-training quantization method using approximate second-order information (Hessian inverse) to minimize quantization error layer-by-layer, achieving near-lossless 4-bit weight compression with calibration data.",
      "prerequisites": [
        "model-quantization",
        "block-wise-quantization"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "awq-quantization",
      "name": "Activation-Aware Weight Quantization",
      "description": "A quantization technique that identifies and preserves salient weight channels (those corresponding to large activation magnitudes) at higher precision, achieving better accuracy than uniform quantization without calibration data at inference time.",
      "prerequisites": [
        "model-quantization",
        "activation-sparsity",
        "block-wise-quantization"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "qlora",
      "name": "QLoRA",
      "description": "Quantized Low-Rank Adaptation: fine-tuning with a 4-bit quantized frozen base model plus FP16 LoRA adapters, enabling fine-tuning of 65B+ parameter models on a single 48 GB GPU by separating static (quantized) and trainable (full-precision) memory.",
      "prerequisites": [
        "low-rank-adaptation",
        "block-wise-quantization"
      ],
      "tracks": [
        "cloud",
        "edge"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "source_domains": [
        "training"
      ],
      "question_count": 0
    },
    {
      "id": "rlhf-infrastructure",
      "name": "RLHF Infrastructure",
      "description": "The systems architecture for Reinforcement Learning from Human Feedback: concurrent serving of policy, reference, reward, and critic models during PPO training, requiring 4x the memory of standard fine-tuning and careful orchestration of generation and training phases.",
      "prerequisites": [
        "capacity-planning",
        "fully-sharded-data-parallel",
        "model-parallelism"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_distributed_training"
      ],
      "source_domains": [
        "distributed_training"
      ],
      "question_count": 0
    },
    {
      "id": "container-gpu-scheduling",
      "name": "Container GPU Scheduling",
      "description": "Kubernetes-native GPU resource management including device plugins, topology-aware scheduling, MIG partitioning, and gang scheduling for multi-GPU jobs, determining utilization efficiency of shared GPU clusters.",
      "prerequisites": [
        "gpu-virtualization",
        "bin-packing-heuristics",
        "topology-aware-placement"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fleet_orchestration"
      ],
      "source_domains": [
        "fleet_orchestration"
      ],
      "question_count": 0
    },
    {
      "id": "inference-accelerator-selection",
      "name": "Inference Accelerator Selection",
      "description": "Systematic evaluation of custom silicon (TPU, Inferentia, Groq, Cerebras) versus GPUs for inference workloads, considering batch size sensitivity, model architecture constraints, cost per token, and vendor lock-in trade-offs.",
      "prerequisites": [
        "total-cost-of-ownership",
        "cost-per-inference",
        "accelerator-spectrum"
      ],
      "tracks": [
        "cloud",
        "edge"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "source_domains": [
        "compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "benchmarking-granularity",
      "name": "Benchmarking Granularity",
      "description": "The structural level at which a system's performance is measured, ranging from isolated hardware operations (micro-benchmarks) to complete application workflows (end-to-end benchmarks).",
      "prerequisites": [
        "system-benchmarking"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_benchmarking"
      ],
      "question_count": 0
    },
    {
      "id": "benchmark-harness",
      "name": "Benchmark Harness",
      "description": "The standardized software infrastructure used to orchestrate benchmark execution, ensure consistent timing, inject data, and validate output correctness across diverse hardware architectures.",
      "prerequisites": [
        "system-benchmarking"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_benchmarking"
      ],
      "question_count": 0
    },
    {
      "id": "benchmark-run-rules",
      "name": "Benchmark Run Rules",
      "description": "The strict constraints governing benchmark execution, including allowable hyperparameter tuning, initialization procedures, and timing definitions, to ensure fair and apples-to-apples comparisons between competing systems.",
      "prerequisites": [
        "mlperf"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_benchmarking"
      ],
      "question_count": 0
    },
    {
      "id": "information-entropy",
      "name": "Information Entropy",
      "description": "A theoretical measure of unpredictability or average information content in a dataset, used to evaluate data diversity and theoretical bounds for compression and transmission.",
      "prerequisites": [
        "data-gravity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_data_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "data-quality-as-code",
      "name": "Data Quality as Code",
      "description": "The engineering practice of codifying data validation rules, schema constraints, and statistical bounds in version-controlled configurations to automate pipeline monitoring.",
      "prerequisites": [
        "schema-validation"
      ],
      "tracks": [
        "cloud",
        "edge"
      ],
      "source_chapters": [
        "vol1_data_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "label-consensus",
      "name": "Label Consensus",
      "description": "A rigorous methodology for aggregating multiple human annotations to ensure high ground-truth quality, often leveraging statistical metrics to resolve ambiguities and filter low-quality annotators.",
      "prerequisites": [
        "dataset-compilation"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_data_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "ai-assisted-labeling",
      "name": "AI-Assisted Labeling",
      "description": "The strategic use of heuristic models or active learning to pre-annotate data or selectively identify the most informative samples, drastically scaling the throughput of human labeling teams.",
      "prerequisites": [
        "dataset-compilation"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_data_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "energy-roofline-model-arithmetic-intensity",
      "name": "Arithmetic Intensity Fundamentals",
      "description": "Understanding, defining, and optimizing arithmetic intensity, the fundamental x-axis metric of the Energy Roofline Model.",
      "prerequisites": [
        "arithmetic-intensity",
        "energy-per-byte",
        "energy-per-operation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "energy-roofline-model-ridge-point",
      "name": "Roofline Ridge Point Analysis",
      "description": "Calculating the ridge point and determining whether operations are compute-bound or memory-bound based on the hardware profile.",
      "prerequisites": [
        "energy-roofline-model-arithmetic-intensity"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "energy-roofline-model-hardware",
      "name": "Hardware-Specific Roofline Modeling",
      "description": "Applying the Energy Roofline Model to analyze specific hardware architectures, from microcontrollers to datacenter GPUs.",
      "prerequisites": [
        "energy-roofline-model-ridge-point"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "energy-roofline-model-workloads",
      "name": "Workload Roofline Optimization",
      "description": "Evaluating specific neural network operations and phases (like LLM generation or depthwise convolutions) against the roofline.",
      "prerequisites": [
        "energy-roofline-model-ridge-point"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "federated-learning-foundations",
      "name": "Federated Learning Foundations",
      "description": "Core principles, architecture, and the fundamental privacy promise of training models across decentralized edge nodes.",
      "prerequisites": [
        "on-device-learning"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_workflow",
        "vol2_edge_intelligence",
        "vol2_security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "federated-learning-economics",
      "name": "Economics of Federated Learning",
      "description": "Financial implications, Total Cost of Ownership (TCO), and evaluating the trade-offs between centralized data infrastructure and federated networks.",
      "prerequisites": [
        "federated-learning-foundations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_responsible_engr",
        "vol2_edge_intelligence"
      ],
      "question_count": 0
    },
    {
      "id": "federated-learning-power",
      "name": "Power & Battery Constraints",
      "description": "Managing the significant battery drain, energy taxes, and power state limitations (active vs. sleep) during on-device training rounds.",
      "prerequisites": [
        "federated-learning-foundations"
      ],
      "tracks": [
        "mobile",
        "tinyml",
        "edge"
      ],
      "source_chapters": [
        "vol2_edge_intelligence",
        "vol2_robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "federated-learning-communication",
      "name": "Communication & Networking",
      "description": "Overcoming the communication tax, cellular data costs, latency, and fleet synchronization challenges inherent in distributed node updates.",
      "prerequisites": [
        "federated-learning-foundations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol2_edge_intelligence",
        "vol2_robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "federated-learning-privacy",
      "name": "Privacy & Security",
      "description": "Security guarantees, breach prevention, and managing the tension between user privacy constraints and model performance.",
      "prerequisites": [
        "federated-learning-foundations"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol1_responsible_engr",
        "vol2_security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "gpu-memory-hierarchy-architecture",
      "name": "Memory Architecture & Characteristics",
      "description": "The fundamental differences between memory tiers, such as SRAM, DRAM, and HBM, and their structural properties.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "gpu-memory-hierarchy-capacity",
      "name": "Memory Capacity & Footprint",
      "description": "Managing the memory footprint of model weights, activations, and optimizer states across different deployment constraints.",
      "prerequisites": [
        "gpu-memory-hierarchy-architecture"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "gpu-memory-hierarchy-bandwidth-latency",
      "name": "Memory Bandwidth & Latency",
      "description": "Analyzing data movement times, latency bottlenecks, and bandwidth requirements for inference and training.",
      "prerequisites": [
        "gpu-memory-hierarchy-architecture"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "gpu-memory-hierarchy-optimization",
      "name": "Memory Access Optimization",
      "description": "Techniques to optimize memory access patterns, including the Roofline model, cache thrashing mitigation, and memory recomputation.",
      "prerequisites": [
        "gpu-memory-hierarchy-capacity",
        "gpu-memory-hierarchy-bandwidth-latency"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "application-embedded-serving-memory-footprint",
      "name": "Memory Footprint and Budgeting",
      "description": "Managing strict memory constraints and minimizing the RAM footprint of models running directly within the host application process.",
      "prerequisites": [],
      "tracks": [
        "mobile",
        "edge",
        "cloud"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "application-embedded-serving-edge-devices",
      "name": "On-Device and Edge Integration",
      "description": "Architecting models to run embedded within mobile, automotive, and microcontroller applications, focusing on local execution and priority inversion.",
      "prerequisites": [
        "application-embedded-serving-memory-footprint"
      ],
      "tracks": [
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "application-embedded-serving-batching-latency",
      "name": "Batching and Latency Optimization",
      "description": "Techniques for batching requests, managing continuous batching, and optimizing latency bounds when inference is embedded tightly with application logic.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "application-embedded-serving-cloud-scale",
      "name": "Large-Scale Embedded Architecture",
      "description": "Design patterns for embedding inference in large-scale cloud services, covering recommendation systems, candidate generation, and production A/B testing.",
      "prerequisites": [
        "application-embedded-serving-batching-latency",
        "application-embedded-serving-memory-footprint"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "application-embedded-serving-ops-hardware",
      "name": "Hardware Utilization and Operations",
      "description": "Operational challenges, data drift, deployment anomalies, and hardware-specific performance limits encountered during embedded serving.",
      "prerequisites": [
        "application-embedded-serving-cloud-scale"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "micro-benchmarking",
      "name": "Micro-Benchmarking",
      "description": "The isolation and measurement of specific, low-level system components or individual operations, such as matrix multiplications or memory transfers, to identify granular performance bottlenecks.",
      "prerequisites": [
        "system-benchmarking"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_benchmarking"
      ],
      "question_count": 0
    },
    {
      "id": "time-to-train",
      "name": "Time-to-Train",
      "description": "The wall-clock time required to train a machine learning model to a predetermined, statistically verified target accuracy, serving as the primary metric for comprehensive ML training benchmarks.",
      "prerequisites": [
        "system-benchmarking",
        "mlperf"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_benchmarking"
      ],
      "question_count": 0
    },
    {
      "id": "inter-annotator-agreement",
      "name": "Inter-Annotator Agreement",
      "description": "Quantitative statistical metrics used to measure the level of consensus and consistency among multiple independent human labelers to establish reliable ground truth.",
      "prerequisites": [
        "dataset-compilation"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_data_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "micro-runtimes",
      "name": "Micro-Runtimes",
      "description": "Specialized framework execution environments designed for highly constrained edge devices, typically operating without dynamic memory allocation and minimizing binary footprint.",
      "prerequisites": [
        "static-computation-graphs",
        "tensor-abstraction"
      ],
      "tracks": [
        "edge",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "neural-module-abstraction",
      "name": "Neural Module Abstraction",
      "description": "An object-oriented design pattern that encapsulates parameters, state, and forward-pass logic into composable blocks, enabling automatic parameter discovery and hierarchical serialization.",
      "prerequisites": [
        "tensor-abstraction",
        "computational-graph"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "mode-dependent-behavior",
      "name": "Mode-Dependent Behavior",
      "description": "The framework mechanism that dynamically alters the computational graph of certain layers (like Dropout or BatchNorm) based on whether the model is in a training or evaluation state.",
      "prerequisites": [
        "computational-graph"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "functional-transformations",
      "name": "Functional Transformations",
      "description": "A framework paradigm operating on pure functions without side effects, enabling composable operations like auto-vectorization, parallelization, and JIT compilation (exemplified by JAX).",
      "prerequisites": [
        "automatic-differentiation",
        "computational-graph"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_frameworks"
      ],
      "question_count": 0
    },
    {
      "id": "domain-specific-architecture",
      "name": "Domain-Specific Architecture (DSA)",
      "description": "A class of microprocessors customized to efficiently execute a specific domain of workloads, such as tensor operations, trading general-purpose flexibility for massive gains in performance and energy efficiency.",
      "prerequisites": [
        "hardware-software-co-design"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "memory-tiling",
      "name": "Memory Tiling",
      "description": "An algorithmic and hardware-mapping optimization that partitions large matrix or tensor operations into smaller sub-blocks (tiles) so the working set fits entirely within fast, on-chip SRAM, minimizing expensive trips to main memory.",
      "prerequisites": [
        "sram",
        "matrix-operations",
        "memory-wall"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "hardware-balance",
      "name": "Hardware Balance",
      "description": "The ratio of an accelerator's peak computational throughput (FLOPs) to its peak memory bandwidth (Bytes/s), defining the critical 'ridge point' where a workload shifts from being memory-bound to compute-bound.",
      "prerequisites": [
        "memory-bandwidth",
        "roofline-model"
      ],
      "tracks": [
        "cloud",
        "edge"
      ],
      "source_chapters": [
        "vol1_hw_acceleration"
      ],
      "question_count": 0
    },
    {
      "id": "ml-test-score",
      "name": "ML Test Score",
      "description": "A quantitative rubric for assessing the operational readiness and technical debt of a machine learning system across data, model, infrastructure, and monitoring dimensions.",
      "prerequisites": [
        "silent-model-failures",
        "boundary-erosion"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_ml_ops"
      ],
      "question_count": 0
    },
    {
      "id": "mlops-maturity-model",
      "name": "MLOps Maturity Model",
      "description": "A framework defining the evolutionary stages of ML operations, from manual, siloed execution (Level 0) to fully automated CI/CD pipelines with continuous training (Level 2).",
      "prerequisites": [
        "model-registry",
        "continuous-training-pipelines"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_ml_ops"
      ],
      "question_count": 0
    },
    {
      "id": "pipeline-jungles",
      "name": "Pipeline Jungles",
      "description": "A form of architectural debt where data preparation and model training workflows become an untestable, monolithic tangle of specialized scripts, intermediate files, and rigid dependencies.",
      "prerequisites": [
        "boundary-erosion"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_ml_ops"
      ],
      "question_count": 0
    },
    {
      "id": "glue-code-antipattern",
      "name": "Glue Code Anti-Pattern",
      "description": "A structural liability where a disproportionate amount of custom integration code is written to adapt generic, off-the-shelf ML packages to specific business domains, locking the system into rigid designs.",
      "prerequisites": [
        "boundary-erosion"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_ml_ops"
      ],
      "question_count": 0
    },
    {
      "id": "undeclared-data-dependencies",
      "name": "Undeclared Data Dependencies",
      "description": "Unintended or implicit linkages between ML models and upstream data sources that can silently degrade model accuracy when the upstream data schema or distribution changes.",
      "prerequisites": [
        "training-serving-skew",
        "silent-model-failures"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_ml_ops"
      ],
      "question_count": 0
    },
    {
      "id": "online-evaluation-strategies",
      "name": "Online Evaluation Strategies",
      "description": "Methodologies such as shadow deployments, A/B testing, and canary releases used to safely validate ML model performance and system behavior in live production environments.",
      "prerequisites": [
        "distribution-shift"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol1_ml_workflow"
      ],
      "question_count": 0
    },
    {
      "id": "system-artifact-provenance",
      "name": "System Artifact Provenance",
      "description": "The rigorous versioning and tracking of all coupled components—including data splits, code, environment configurations, and model weights—required to ensure deterministic reproducibility of an ML system.",
      "prerequisites": [
        "data-lineage"
      ],
      "tracks": [
        "cloud",
        "edge"
      ],
      "source_chapters": [
        "vol1_ml_workflow"
      ],
      "question_count": 0
    },
    {
      "id": "deployment-pareto-optimization",
      "name": "Deployment Pareto Optimization",
      "description": "The systematic balancing of a model's predictive accuracy against operational deployment constraints such as latency, memory footprint, and power consumption.",
      "prerequisites": [
        "constraint-propagation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_ml_workflow"
      ],
      "question_count": 0
    },
    {
      "id": "littles-law",
      "name": "Little's Law",
      "description": "A queuing theory theorem defining the relationship between concurrency, arrival rate, and latency (L = λW). It is essential for determining maximum safe concurrency and capacity planning.",
      "prerequisites": [
        "capacity-planning",
        "throughput",
        "latency",
        "kv-cache"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_inference",
        "vol1_model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "request-pipelining",
      "name": "Request Pipelining",
      "description": "The technique of overlapping different execution stages (e.g., CPU preprocessing, GPU execution, network I/O) across multiple concurrent requests to maximize hardware utilization and system throughput.",
      "prerequisites": [
        "throughput",
        "inference-pipeline"
      ],
      "tracks": [
        "cloud",
        "edge"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "tail-tolerant-serving",
      "name": "Tail-Tolerant Serving",
      "description": "Architectural techniques such as hedged requests, tied requests, and microservice isolation designed to mitigate tail latency and ensure predictable response times in distributed serving systems.",
      "prerequisites": [
        "tail-latency",
        "load-balancer"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "cold-start",
      "name": "Cold Start Penalty",
      "description": "The initial, significantly higher latency experienced when a model is first invoked, caused by loading weights into memory, initializing the runtime, and executing first-pass graph compilation.",
      "prerequisites": [
        "latency",
        "inference-server"
      ],
      "tracks": [
        "cloud",
        "edge"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "multi-model-serving",
      "name": "Multi-Model Serving",
      "description": "An architecture where a single inference server process or hardware accelerator simultaneously hosts and routes traffic to multiple distinct ML models to maximize resource utilization.",
      "prerequisites": [
        "inference-server"
      ],
      "tracks": [
        "cloud",
        "edge"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "model-caching",
      "name": "Model Caching Infrastructure",
      "description": "Systems that manage memory hierarchy by keeping frequently accessed models in fast memory (e.g., GPU VRAM) while dynamically loading or evicting less active models from slower storage.",
      "prerequisites": [
        "multi-model-serving",
        "inference-server"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_model_serving"
      ],
      "question_count": 0
    },
    {
      "id": "positional-encoding",
      "name": "Positional Encoding",
      "description": "A structural mechanism to inject sequence order information into parallel processing architectures like Transformers, often implemented via trigonometric functions or learned embeddings.",
      "prerequisites": [
        "transformer-architecture"
      ],
      "tracks": [
        "cloud",
        "edge"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "receptive-field",
      "name": "Receptive Field",
      "description": "The specific region of the input space that affects a given activation in a convolutional neural network layer, dictating how much spatial context the model can capture at that depth.",
      "prerequisites": [
        "convolutional-neural-network"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_architectures"
      ],
      "question_count": 0
    },
    {
      "id": "hierarchical-feature-learning",
      "name": "Hierarchical Feature Learning",
      "description": "The mechanism by which deep networks progressively compose simple, low-level patterns into complex, high-level semantic abstractions, justifying the computational cost of additional layers.",
      "prerequisites": [
        "dense-connectivity-scaling"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_computation"
      ],
      "question_count": 0
    },
    {
      "id": "end-to-end-differentiability",
      "name": "End-to-End Differentiability",
      "description": "The architectural requirement that all components in a machine learning pipeline must support continuous gradient flow, enabling automated pattern discovery and eliminating the feature engineering bottleneck.",
      "prerequisites": [
        "credit-assignment-problem"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_computation"
      ],
      "question_count": 0
    },
    {
      "id": "backpropagation-compute-cost",
      "name": "Backpropagation Compute Cost",
      "description": "The system reality that the backward pass requires roughly twice the computational operations (FLOPs) of the forward pass, heavily impacting training infrastructure requirements.",
      "prerequisites": [
        "multiply-accumulate-operations",
        "credit-assignment-problem"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_nn_computation"
      ],
      "question_count": 0
    },
    {
      "id": "inference-pipeline-latency",
      "name": "Inference Pipeline Latency",
      "description": "The critical end-to-end execution time required for the forward pass during production deployment, determining the system's ability to meet real-time operational SLOs.",
      "prerequisites": [
        "matrix-multiplication-workloads"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol1_nn_computation"
      ],
      "question_count": 0
    },
    {
      "id": "runaway-feedback-loops",
      "name": "Runaway Feedback Loops",
      "description": "A silent failure mode where a model's actions in production continuously bias the subsequent data collected for retraining, amplifying errors or inequalities over time.",
      "prerequisites": [
        "distribution-shift",
        "alignment-gap"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_responsible_engr"
      ],
      "question_count": 0
    },
    {
      "id": "model-fallback-strategies",
      "name": "Model Fallback Strategies",
      "description": "Architectural patterns for gracefully degrading system functionality by reverting to simpler heuristics, cached responses, or human overrides during anomalies.",
      "prerequisites": [
        "distribution-shift"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol1_responsible_engr"
      ],
      "question_count": 0
    },
    {
      "id": "backpropagation-compute-ratio",
      "name": "Backpropagation Compute Ratio",
      "description": "The fundamental heuristic that the backward pass requires approximately twice the FLOPs of the forward pass, used to estimate total mathematical operations per training step.",
      "prerequisites": [
        "computational-intensity"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "question_count": 0
    },
    {
      "id": "training-physical-ceiling",
      "name": "Training Physical Ceiling",
      "description": "The ultimate hardware limitations—such as interconnect latency, power density, and thermal dissipation—that bound the maximum effective scale of a synchronous training cluster before efficiency collapses.",
      "prerequisites": [
        "iron-law-of-training",
        "staged-system-pipeline"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol1_training"
      ],
      "question_count": 0
    },
    {
      "id": "logp-model",
      "name": "LogP Model",
      "description": "A formal network communication model characterizing performance via Latency, Overhead, Gap, and Processor parameters, providing deeper insight for scheduling than linear models.",
      "prerequisites": [
        "alpha-beta-performance-model"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_collective_communication"
      ],
      "question_count": 0
    },
    {
      "id": "tree-allreduce",
      "name": "Tree AllReduce",
      "description": "A collective communication algorithm that organizes nodes into a logical tree structure (or double binary tree), significantly reducing latency for small messages compared to ring topologies.",
      "prerequisites": [
        "allreduce"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_collective_communication"
      ],
      "question_count": 0
    },
    {
      "id": "recursive-halving-doubling",
      "name": "Recursive Halving-Doubling",
      "description": "A bandwidth-optimal butterfly communication algorithm that minimizes network steps for exact power-of-two node counts without relying on a physical ring.",
      "prerequisites": [
        "allreduce"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_collective_communication"
      ],
      "question_count": 0
    },
    {
      "id": "in-network-reduction",
      "name": "In-Network Reduction",
      "description": "The practice of offloading collective mathematical operations directly to network switch hardware (e.g., SHARP) to reduce host CPU/GPU overhead and halve network traffic.",
      "prerequisites": [
        "allreduce"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_collective_communication"
      ],
      "question_count": 0
    },
    {
      "id": "topology-aware-routing",
      "name": "Topology-Aware Routing",
      "description": "The strategic scheduling of network data transfers to exploit the underlying physical architecture, avoiding cross-rack or cross-spine bottlenecks in hierarchical deployments.",
      "prerequisites": [
        "hierarchical-allreduce"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_collective_communication"
      ],
      "question_count": 0
    },
    {
      "id": "compression-aware-optimization",
      "name": "Compression-Aware Optimization",
      "description": "Specialized training optimizers (such as 1-bit Adam) engineered to natively process aggressively quantized or sparsified gradients without degrading model convergence.",
      "prerequisites": [
        "gradient-compression"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_collective_communication"
      ],
      "question_count": 0
    },
    {
      "id": "high-speed-gpu-interconnect",
      "name": "High-Speed GPU Interconnect",
      "description": "Proprietary node-level memory fabrics (e.g., NVLink, NVSwitch) that bypass PCIe bottlenecks to enable high-bandwidth, unified memory spaces across multiple GPUs within a single dense server node.",
      "prerequisites": [
        "pcie",
        "memory-wall"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "non-uniform-memory-access",
      "name": "Non-Uniform Memory Access (NUMA)",
      "description": "A memory architecture in dense ML nodes where memory access time depends on the memory location relative to the processor, requiring strict CPU-GPU-NIC affinity rules to avoid cross-socket PCIe bottlenecks.",
      "prerequisites": [
        "pcie"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "direct-storage-access",
      "name": "Direct Storage Access",
      "description": "I/O bypass technologies (like GPUDirect Storage) that allow high-speed NVMe drives to transfer training data directly via PCIe to GPU memory via DMA, eliminating CPU bounce buffers and reducing latency.",
      "prerequisites": [
        "pcie",
        "memory-wall"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "rack-power-density",
      "name": "Rack Power Density",
      "description": "A critical facility-level constraint measuring the total kilowatt (kW) draw per server rack, which dictates the layout, power delivery mechanisms, and necessary cooling paradigms for dense ML clusters.",
      "prerequisites": [
        "thermal-design-power"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "direct-to-chip-liquid-cooling",
      "name": "Direct-to-Chip Liquid Cooling",
      "description": "Advanced thermal management systems required for modern high-TDP compute nodes, where cold plates route liquid coolant directly over silicon packages to prevent thermal throttling.",
      "prerequisites": [
        "thermal-design-power"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "infrastructure-tco",
      "name": "Infrastructure Total Cost of Ownership",
      "description": "Financial modeling for ML hardware lifecycle that balances upfront Capital Expenditure (CapEx) of accelerators against ongoing Operational Expenditure (OpEx) including power, cooling, network, and depreciation.",
      "prerequisites": [
        "thermal-design-power"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_compute_infrastructure"
      ],
      "question_count": 0
    },
    {
      "id": "stale-synchronous-parallel",
      "name": "Stale Synchronous Parallel (SSP)",
      "description": "A distributed synchronization model that bounds the staleness of worker gradients, offering a trade-off between the strict consistency of BSP and the high throughput of ASP.",
      "prerequisites": [
        "data-parallelism"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_distributed_training"
      ],
      "question_count": 0
    },
    {
      "id": "learning-rate-scaling-rules",
      "name": "Learning Rate Scaling Rules",
      "description": "Heuristics, such as the linear scaling rule, required to adjust learning rates proportionally when scaling global batch sizes to maintain stable convergence.",
      "prerequisites": [
        "data-parallelism"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_distributed_training"
      ],
      "question_count": 0
    },
    {
      "id": "critical-batch-size",
      "name": "Critical Batch Size",
      "description": "The maximum effective batch size beyond which adding more parallel workers yields diminishing returns in convergence time and computational efficiency.",
      "prerequisites": [
        "data-parallelism"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_distributed_training"
      ],
      "question_count": 0
    },
    {
      "id": "parameter-server-architecture",
      "name": "Parameter Server Architecture",
      "description": "A distributed computing paradigm where model parameters are sharded across dedicated servers, and worker nodes fetch weights and push gradients.",
      "prerequisites": [
        "data-parallelism"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_distributed_training"
      ],
      "question_count": 0
    },
    {
      "id": "expert-parallelism",
      "name": "Expert Parallelism",
      "description": "A parallelism strategy designed for Mixture of Experts (MoE) models, routing tokens across devices to specific expert subnetworks to scale model capacity efficiently.",
      "prerequisites": [
        "model-flops-utilization"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_distributed_training"
      ],
      "question_count": 0
    },
    {
      "id": "fp8-training",
      "name": "FP8 Training",
      "description": "The adoption of 8-bit floating-point formats (such as E4M3 and E5M2) to accelerate training compute and alleviate memory bandwidth bottlenecks while maintaining numerical stability.",
      "prerequisites": [
        "compute-bound",
        "communication-bound"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_distributed_training"
      ],
      "question_count": 0
    },
    {
      "id": "bathtub-curve",
      "name": "Bathtub Curve",
      "description": "A classic reliability model that describes the hazard rate of hardware components over time, characterized by early 'infant mortality' failures, a long period of constant random failures, and a final 'wear-out' phase.",
      "prerequisites": [
        "failure-rate"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "mean-time-between-failures",
      "name": "Mean Time Between Failures (MTBF)",
      "description": "A key reliability metric representing the expected elapsed time between inherent failures of a system during normal operation, crucial for calculating cluster-level reliability and checkpoint intervals.",
      "prerequisites": [
        "failure-rate"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "permanent-faults",
      "name": "Permanent Faults",
      "description": "Hardware failures that cause irreversible damage to a component (e.g., a burnt-out transistor), requiring physical replacement or permanent software-level rerouting.",
      "prerequisites": [
        "fault-model"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "intermittent-faults",
      "name": "Intermittent Faults",
      "description": "Failures that occur irregularly and occasionally, often caused by marginal hardware components, thermal fluctuations, or loose connections, making them notoriously difficult to detect and reproduce.",
      "prerequisites": [
        "fault-model"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "checkpoint-storm",
      "name": "Checkpoint Storm",
      "description": "A system-wide performance collapse that occurs when thousands of distributed workers simultaneously attempt to write massive state files to a shared file system, overwhelming network and storage bandwidth.",
      "prerequisites": [
        "distributed-checkpointing",
        "checkpointing"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_fault_tolerance"
      ],
      "question_count": 0
    },
    {
      "id": "five-level-networking-model",
      "name": "Five-Level Networking Model",
      "description": "A conceptual framework for understanding ML network fabrics, encompassing wire and link, transport, switch topology, fabric behavior, and cluster design.",
      "prerequisites": [
        "alpha-beta-performance-model"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "link-budget",
      "name": "Link Budget",
      "description": "The rigorous accounting of signal power gains and losses from transmitter to receiver, determining the maximum viable distance and speed for copper or optical interconnects.",
      "prerequisites": [
        "pam4",
        "serializer-deserializer"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "failure-domain",
      "name": "Failure Domain",
      "description": "A logical or physical boundary within a cluster topology, such as a Top-of-Rack switch, where a single component failure causes an outage for all downstream connected resources.",
      "prerequisites": [
        "top-of-rack-switch"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "gradient-bus",
      "name": "Gradient Bus",
      "description": "A dedicated, high-bandwidth backend network fabric optimized specifically for the rapid collective communication of model gradients directly between GPUs.",
      "prerequisites": [
        "allreduce",
        "gpudirect-rdma"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_network_fabrics"
      ],
      "question_count": 0
    },
    {
      "id": "ml-technical-debt",
      "name": "ML Technical Debt",
      "description": "Identifying and mitigating hidden maintenance costs in machine learning platforms, such as pipeline jungles, data dependencies, and configuration sprawl.",
      "prerequisites": [
        "model-registry"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "ml-capacity-planning",
      "name": "ML Capacity Planning",
      "description": "Forecasting and provisioning compute resources and estimating the cost of training and serving to ensure platform economic viability and cost efficiency.",
      "prerequisites": [
        "serving-autoscaling"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "ensemble-management",
      "name": "Ensemble Model Management",
      "description": "Orchestrating the lifecycle, versioning, and deployment of composite multi-model architectures to ensure consistent predictions across dependent models.",
      "prerequisites": [
        "model-registry",
        "dependency-aware-deployment"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "continuous-training-pipeline",
      "name": "Continuous Training Pipeline",
      "description": "Automating the end-to-end retraining process triggered by data drift or schedule, including data ingestion, training, and artifact generation at scale.",
      "prerequisites": [
        "model-registry"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "automated-validation-gates",
      "name": "Automated Validation Gates",
      "description": "Implementing strict programmatic quality checks, such as performance thresholds and fairness constraints, before an ML model artifact is automatically promoted to deployment.",
      "prerequisites": [
        "model-registry"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "hierarchical-monitoring",
      "name": "Hierarchical Monitoring Architecture",
      "description": "Structuring system observability from high-level business metrics down to granular model and system telemetry to effectively isolate failures across massive multi-model fleets.",
      "prerequisites": [
        "telemetry-aggregation"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "alert-fatigue-mitigation",
      "name": "Alert Fatigue Mitigation",
      "description": "Strategies for tuning monitoring thresholds, aggregating alerts, and implementing anomaly detection to prevent operators from ignoring critical system warnings in multi-model environments.",
      "prerequisites": [
        "telemetry-aggregation"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_ops_scale"
      ],
      "question_count": 0
    },
    {
      "id": "neural-architecture-search",
      "name": "Neural Architecture Search",
      "description": "An automated algorithmic process for designing neural network topologies, replacing manual architectural engineering with search over a defined operation space.",
      "prerequisites": [
        "efficiency-frontier"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "hardware-aware-nas",
      "name": "Hardware-Aware NAS",
      "description": "An advanced AutoML strategy that incorporates hardware-specific constraints such as latency, memory footprint, and power consumption directly into the architectural search objective.",
      "prerequisites": [
        "efficiency-frontier",
        "roofline-model"
      ],
      "tracks": [
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "automated-model-compression",
      "name": "Automated Model Compression",
      "description": "The use of search algorithms or reinforcement learning to automatically determine the optimal compression policy (e.g., pruning sparsity or quantization bit-widths) per layer.",
      "prerequisites": [
        "efficiency-frontier",
        "precision-engineering"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_performance_engineering"
      ],
      "question_count": 0
    },
    {
      "id": "system-feedback-loops",
      "name": "System Feedback Loops",
      "description": "The phenomenon where a deployed model's predictions causally influence the generation or selection of future training data, potentially amplifying historical biases or creating runaway filter bubbles.",
      "prerequisites": [
        "distribution-shift"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol2_responsible_ai"
      ],
      "question_count": 0
    },
    {
      "id": "algorithmic-contestability",
      "name": "Algorithmic Contestability",
      "description": "System architecture patterns that allow end-users or human-in-the-loop operators to appeal, review, and systematically override automated algorithmic decisions.",
      "prerequisites": [
        "system-explainability"
      ],
      "tracks": [
        "cloud",
        "edge"
      ],
      "source_chapters": [
        "vol2_responsible_ai"
      ],
      "question_count": 0
    },
    {
      "id": "llm-guardrails",
      "name": "LLM Guardrails",
      "description": "Boundary defense systems using semantic routing, secondary classifier models, or deterministic filters to intercept and block policy-violating prompts or generative outputs.",
      "prerequisites": [
        "adversarial-inputs"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol2_responsible_ai"
      ],
      "question_count": 0
    },
    {
      "id": "differential-privacy-mechanisms",
      "name": "Differential Privacy Mechanisms",
      "description": "Mathematical frameworks, such as DP-SGD, that inject calibrated noise during model training or aggregation to mathematically guarantee that individual data records cannot be reverse-engineered.",
      "prerequisites": [
        "membership-inference-attack"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol2_responsible_ai"
      ],
      "question_count": 0
    },
    {
      "id": "out-of-distribution-detection",
      "name": "Out-of-Distribution Detection",
      "description": "Techniques used during inference to identify and flag inputs that fall significantly outside the model's training data distribution, preventing unpredictable behavior and silent failures.",
      "prerequisites": [
        "distribution-shift"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol2_robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "prompt-injection",
      "name": "Prompt Injection",
      "description": "A form of adversarial attack against generative AI systems where malicious instructions are embedded within user inputs to bypass safety guardrails and hijack the model's intended behavior.",
      "prerequisites": [
        "adversarial-attacks"
      ],
      "tracks": [
        "cloud",
        "edge"
      ],
      "source_chapters": [
        "vol2_robust_ai"
      ],
      "question_count": 0
    },
    {
      "id": "threat-modeling-ml",
      "name": "ML Threat Modeling",
      "description": "A systematic methodology for identifying, enumerating, and prioritizing potential vulnerabilities specific to machine learning workflows and architectures.",
      "prerequisites": [
        "layered-defense"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "adversarial-evasion-attack",
      "name": "Adversarial Evasion Attack",
      "description": "Crafting imperceptible perturbations to input data at inference time designed to deceive a deployed model into making incorrect predictions.",
      "prerequisites": [
        "runtime-input-validation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "llm-prompt-injection",
      "name": "LLM Prompt Injection",
      "description": "Maliciously formatting inputs to a Large Language Model to manipulate its execution flow, bypass safety guardrails, or leak system prompts and sensitive data.",
      "prerequisites": [
        "runtime-input-validation"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "ml-supply-chain-security",
      "name": "ML Supply Chain Security",
      "description": "Securing the provenance and integrity of third-party datasets, pre-trained model weights, and software dependencies to prevent upstream compromise from poisoning the ML system.",
      "prerequisites": [
        "layered-defense"
      ],
      "tracks": [
        "cloud",
        "edge"
      ],
      "source_chapters": [
        "vol2_security_privacy"
      ],
      "question_count": 0
    },
    {
      "id": "water-usage-effectiveness",
      "name": "Water Usage Effectiveness",
      "description": "A facility-level sustainability metric measuring the ratio of total water used by a datacenter to its IT equipment energy consumption, quantifying the environmental cost of cooling high-density AI clusters.",
      "prerequisites": [
        "power-usage-effectiveness"
      ],
      "tracks": [
        "cloud"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "grid-carbon-intensity",
      "name": "Grid Carbon Intensity",
      "description": "A dynamically varying metric that quantifies the greenhouse gas emissions produced per unit of electrical energy consumed from a specific regional power grid, typically measured in gCO2eq/kWh.",
      "prerequisites": [
        "operational-carbon"
      ],
      "tracks": [
        "cloud",
        "edge"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "embodied-carbon-amortization",
      "name": "Embodied Carbon Amortization",
      "description": "The accounting practice and systems engineering strategy of extending hardware lifespans to distribute fixed manufacturing emissions over a longer operational period, reducing the annualized carbon cost per AI workload.",
      "prerequisites": [
        "embodied-carbon",
        "lifecycle-carbon-accounting"
      ],
      "tracks": [
        "cloud",
        "edge",
        "mobile"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "divergent-energy-scaling",
      "name": "Divergent Energy Scaling",
      "description": "The systemic trajectory where the computational and memory demands of state-of-the-art AI models grow exponentially faster than the energy efficiency improvements of underlying hardware, eventually hitting hard physical power delivery constraints.",
      "prerequisites": [
        "cmos-power-equation",
        "power-density-ceiling"
      ],
      "tracks": [
        "cloud",
        "edge"
      ],
      "source_chapters": [
        "vol2_sustainable_ai"
      ],
      "question_count": 0
    },
    {
      "id": "distributed-consensus",
      "name": "Distributed Consensus",
      "description": "Agreement protocols (Paxos, Raft) for maintaining consistency across distributed ML serving and training infrastructure.",
      "prerequisites": [
        "fault-tolerance"
      ],
      "tracks": [
        "cloud",
        "global"
      ],
      "source_chapters": [
        "vol2_distributed_training"
      ],
      "source_domains": [
        "distributed-systems"
      ],
      "question_count": 0
    },
    {
      "id": "streaming-ingestion",
      "name": "Streaming Ingestion",
      "description": "Real-time data ingestion pipelines with message delivery guarantees (at-least-once, exactly-once) for ML training and serving.",
      "prerequisites": [
        "data-pipeline-equation"
      ],
      "tracks": [
        "cloud",
        "global"
      ],
      "source_chapters": [
        "vol2_data_engineering"
      ],
      "source_domains": [
        "data-engineering"
      ],
      "question_count": 0
    },
    {
      "id": "responsible-ai",
      "name": "Responsible AI",
      "description": "Regulatory compliance (EU AI Act, GDPR), data sovereignty, bias auditing, and governance frameworks for ML systems deployment.",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "global"
      ],
      "source_chapters": [
        "vol2_responsible_ai"
      ],
      "source_domains": [
        "governance"
      ],
      "question_count": 0
    },
    {
      "id": "precision-quantization-effects",
      "name": "Precision Quantization Effects",
      "description": "Auto-generated concept for precision-quantization-effects",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "interconnect-topology",
      "name": "Interconnect Topology",
      "description": "Auto-generated concept for interconnect-topology",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "parallelism-strategies",
      "name": "Parallelism Strategies",
      "description": "Auto-generated concept for parallelism-strategies",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "kv-cache-memory-management",
      "name": "Kv Cache Memory Management",
      "description": "Auto-generated concept for kv-cache-memory-management",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "kernel-fusion-optimization",
      "name": "Kernel Fusion Optimization",
      "description": "Auto-generated concept for kernel-fusion-optimization",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "training-memory-budget",
      "name": "Training Memory Budget",
      "description": "Auto-generated concept for training-memory-budget",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "allreduce-communication-patterns",
      "name": "Allreduce Communication Patterns",
      "description": "Auto-generated concept for allreduce-communication-patterns",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "inference-optimization",
      "name": "Inference Optimization",
      "description": "Auto-generated concept for inference-optimization",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "energy-efficiency-metrics",
      "name": "Energy Efficiency Metrics",
      "description": "Auto-generated concept for energy-efficiency-metrics",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "data-loading-pipeline",
      "name": "Data Loading Pipeline",
      "description": "Auto-generated concept for data-loading-pipeline",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "latency-throughput-tradeoff",
      "name": "Latency Throughput Tradeoff",
      "description": "Auto-generated concept for latency-throughput-tradeoff",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "checkpoint-restart",
      "name": "Checkpoint Restart",
      "description": "Auto-generated concept for checkpoint-restart",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "fault-tolerance-mechanisms",
      "name": "Fault Tolerance Mechanisms",
      "description": "Auto-generated concept for fault-tolerance-mechanisms",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "training-monitoring",
      "name": "Training Monitoring",
      "description": "Auto-generated concept for training-monitoring",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "operational-intensity",
      "name": "Operational Intensity",
      "description": "Concept for Operational Intensity",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "memory-bandwidth-utilization",
      "name": "Memory Bandwidth Utilization",
      "description": "Concept for Memory Bandwidth Utilization",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "token-generation-throughput",
      "name": "Token Generation Throughput",
      "description": "Concept for Token Generation Throughput",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "tensor-parallelism-memory-footprint",
      "name": "Tensor Parallelism Memory Footprint",
      "description": "Concept for Tensor Parallelism Memory Footprint",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "pipeline-bubble-delay",
      "name": "Pipeline Bubble Delay",
      "description": "Concept for Pipeline Bubble Delay",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "compute-utilization-factor",
      "name": "Compute Utilization Factor",
      "description": "Concept for Compute Utilization Factor",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "performance-per-watt",
      "name": "Performance Per Watt",
      "description": "Concept for Performance Per Watt",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "kv-cache-capacity",
      "name": "Kv Cache Capacity",
      "description": "Concept for KV Cache Capacity",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "interconnect-bottleneck",
      "name": "Interconnect Bottleneck",
      "description": "Concept for Interconnect Bottleneck",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "sram-constraints",
      "name": "Sram Constraints",
      "description": "Concept for SRAM Constraints",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "sub-byte-quantization",
      "name": "Sub Byte Quantization",
      "description": "Concept for Sub byte Quantization",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "pipeline-parallelism-memory-distribution",
      "name": "Pipeline Parallelism Memory Distribution",
      "description": "Concept for Pipeline Parallelism Memory Distribution",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "ring-all-reduce-bandwidth",
      "name": "Ring All Reduce Bandwidth",
      "description": "Concept for Ring All Reduce Bandwidth",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "cost-per-token",
      "name": "Cost Per Token",
      "description": "Concept for Cost per Token",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "datacenter-pue",
      "name": "Datacenter Pue",
      "description": "Concept for Datacenter PUE",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "maximum-batch-size",
      "name": "Maximum Batch Size",
      "description": "Concept for Maximum Batch Size",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "attention-operational-intensity",
      "name": "Attention Operational Intensity",
      "description": "Concept for Attention Operational Intensity",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "kv-cache-sizing",
      "name": "Kv Cache Sizing",
      "description": "Concept for KV Cache Sizing",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "mixed-precision-memory-footprint",
      "name": "Mixed Precision Memory Footprint",
      "description": "Concept for Mixed Precision Memory Footprint",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "bandwidth-delay-product",
      "name": "Bandwidth Delay Product",
      "description": "Concept for Bandwidth Delay Product",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "simd-alignment",
      "name": "Simd Alignment",
      "description": "Concept for SIMD Alignment",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "dispatch-overhead",
      "name": "Dispatch Overhead",
      "description": "Concept for Dispatch Overhead",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "data-reuse",
      "name": "Data Reuse",
      "description": "Concept for Data Reuse",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "sram-spilling",
      "name": "Sram Spilling",
      "description": "Concept for SRAM Spilling",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "weight-streaming",
      "name": "Weight Streaming",
      "description": "Concept for Weight Streaming",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "double-buffering",
      "name": "Double Buffering",
      "description": "Concept for Double Buffering",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "tiling-strategies",
      "name": "Tiling Strategies",
      "description": "Concept for Tiling Strategies",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "fake-quantization",
      "name": "Fake Quantization",
      "description": "Concept for Fake Quantization",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "asymmetric-overhead",
      "name": "Asymmetric Overhead",
      "description": "Concept for Asymmetric Overhead",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "mixed-precision-overhead",
      "name": "Mixed Precision Overhead",
      "description": "Concept for Mixed precision Overhead",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "per-channel-scaling",
      "name": "Per Channel Scaling",
      "description": "Concept for Per channel Scaling",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "initialization-overhead",
      "name": "Initialization Overhead",
      "description": "Concept for Initialization Overhead",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "binary-bloat",
      "name": "Binary Bloat",
      "description": "Concept for Binary Bloat",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "memory-fragmentation",
      "name": "Memory Fragmentation",
      "description": "Concept for Memory Fragmentation",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "ota-delta-updates",
      "name": "Ota Delta Updates",
      "description": "Concept for OTA Delta Updates",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "dram-energy-cost",
      "name": "Dram Energy Cost",
      "description": "Concept for DRAM Energy Cost",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "sleep-states",
      "name": "Sleep States",
      "description": "Concept for Sleep States",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "leakage-power",
      "name": "Leakage Power",
      "description": "Concept for Leakage Power",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "pipeline-energy",
      "name": "Pipeline Energy",
      "description": "Concept for Pipeline Energy",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "redundant-inference",
      "name": "Redundant Inference",
      "description": "Concept for Redundant Inference",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "skin-temperature-limit",
      "name": "Skin Temperature Limit",
      "description": "Concept for Skin Temperature Limit",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "thermal-capacitance",
      "name": "Thermal Capacitance",
      "description": "Concept for Thermal Capacitance",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "thermal-shutdown",
      "name": "Thermal Shutdown",
      "description": "Concept for Thermal Shutdown",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "memory-contention",
      "name": "Memory Contention",
      "description": "Concept for Memory Contention",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "thread-starvation",
      "name": "Thread Starvation",
      "description": "Concept for Thread Starvation",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "memory-layout-nchw-nhwc",
      "name": "Memory Layout (Nchw/Nhwc)",
      "description": "Concept for Memory Layout (NCHW/NHWC)",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "compute-vs-memory-bound",
      "name": "Compute Vs Memory Bound",
      "description": "Concept for Compute vs Memory Bound",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "memory-hierarchies",
      "name": "Memory Hierarchies",
      "description": "Concept for Memory Hierarchies",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "scaling-laws",
      "name": "Scaling Laws",
      "description": "Concept for Scaling Laws",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "arithmetic-intensity-shift",
      "name": "Arithmetic Intensity Shift",
      "description": "Concept for Arithmetic Intensity Shift",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "instruction-overhead",
      "name": "Instruction Overhead",
      "description": "Concept for Instruction Overhead",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "static-vs-dynamic",
      "name": "Static Vs Dynamic",
      "description": "Concept for Static vs Dynamic",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "pruning-vs-quantization",
      "name": "Pruning Vs Quantization",
      "description": "Concept for Pruning vs Quantization",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "batching-latency",
      "name": "Batching Latency",
      "description": "Concept for Batching Latency",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "throughput-vs-latency",
      "name": "Throughput Vs Latency",
      "description": "Concept for Throughput vs Latency",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "hardware-offloading",
      "name": "Hardware Offloading",
      "description": "Concept for Hardware Offloading",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "duty-cycling",
      "name": "Duty Cycling",
      "description": "Concept for Duty Cycling",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "power-profiles",
      "name": "Power Profiles",
      "description": "Concept for Power Profiles",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "dvfs-scaling",
      "name": "Dvfs Scaling",
      "description": "Concept for DVFS Scaling",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "unified-memory",
      "name": "Unified Memory",
      "description": "Concept for Unified Memory",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "pcie-bottlenecks",
      "name": "Pcie Bottlenecks",
      "description": "Concept for PCIe Bottlenecks",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "software-caching",
      "name": "Software Caching",
      "description": "Concept for Software Caching",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "high-availability-models",
      "name": "High Availability Models",
      "description": "Concept for High Availability Models",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "state-checkpointing",
      "name": "State Checkpointing",
      "description": "Concept for State Checkpointing",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "batch-size-sla",
      "name": "Batch Size Sla",
      "description": "Concept for Batch Size SLA",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "multi-process-service",
      "name": "Multi Process Service",
      "description": "Concept for Multi Process Service",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "edge-vs-cloud",
      "name": "Edge Vs Cloud",
      "description": "Concept for Edge vs Cloud",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "communication-vs-compute",
      "name": "Communication Vs Compute",
      "description": "Concept for Communication vs Compute",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "split-inference",
      "name": "Split Inference",
      "description": "Concept for Split Inference",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "3d-parallelism-topology",
      "name": "3D Parallelism Topology",
      "description": "Concept for 3D Parallelism Topology",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "asynchronous-checkpointing",
      "name": "Asynchronous Checkpointing",
      "description": "Concept for Asynchronous Checkpointing",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "tiered-memory-management",
      "name": "Tiered Memory Management",
      "description": "Concept for Tiered Memory Management",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "dynamic-shape-compilation",
      "name": "Dynamic Shape Compilation",
      "description": "Concept for Dynamic Shape Compilation",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "moe-network-topology",
      "name": "Moe Network Topology",
      "description": "Concept for MoE Network Topology",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "pagedattention-telemetry",
      "name": "Pagedattention Telemetry",
      "description": "Concept for PagedAttention Telemetry",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "multi-modal-decoupling",
      "name": "Multi Modal Decoupling",
      "description": "Concept for Multi Modal Decoupling",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "elastic-parameter-server",
      "name": "Elastic Parameter Server",
      "description": "Concept for Elastic Parameter Server",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "edge-quantization-and-fusion",
      "name": "Edge Quantization And Fusion",
      "description": "Concept for Edge Quantization and Fusion",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "federated-lora-training",
      "name": "Federated Lora Training",
      "description": "Concept for Federated LoRA Training",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "zero-copy-serverless-offload",
      "name": "Zero Copy Serverless Offload",
      "description": "Concept for Zero Copy Serverless Offload",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "silent-data-corruption-telemetry",
      "name": "Silent Data Corruption Telemetry",
      "description": "Concept for Silent Data Corruption Telemetry",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "sequence-parallelism",
      "name": "Sequence Parallelism",
      "description": "Concept for Sequence Parallelism",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "resilient-all-reduce-topology",
      "name": "Resilient All Reduce Topology",
      "description": "Concept for Resilient All Reduce Topology",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "auto-tuning-compilers",
      "name": "Auto Tuning Compilers",
      "description": "Concept for Auto Tuning Compilers",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "finops-asymmetric-scheduling",
      "name": "Finops Asymmetric Scheduling",
      "description": "Concept for FinOps Asymmetric Scheduling",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "p99-cuda-event-tracing",
      "name": "P99 Cuda Event Tracing",
      "description": "Concept for P99 CUDA Event Tracing",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "selective-activation-recomputation",
      "name": "Selective Activation Recomputation",
      "description": "Concept for Selective Activation Recomputation",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "pipeline-parallel-serving",
      "name": "Pipeline Parallel Serving",
      "description": "Concept for Pipeline Parallel Serving",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "copy-on-write-shared-caching",
      "name": "Copy On Write Shared Caching",
      "description": "Concept for Copy On Write Shared Caching",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "kv-cache-resiliency",
      "name": "Kv Cache Resiliency",
      "description": "Concept for KV Cache Resiliency",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "asymmetric-pipeline-parallelism",
      "name": "Asymmetric Pipeline Parallelism",
      "description": "Concept for Asymmetric Pipeline Parallelism",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "tiered-traffic-routing",
      "name": "Tiered Traffic Routing",
      "description": "Concept for Tiered Traffic Routing",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "hbm-generation-comparison",
      "name": "Hbm Generation Comparison",
      "description": "Concept for HBM generation comparison",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "legacy-gpu-memory-specs",
      "name": "Legacy Gpu Memory Specs",
      "description": "Concept for Legacy GPU memory specs",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "pcie-bandwidth-recall",
      "name": "Pcie Bandwidth Recall",
      "description": "Concept for PCIe Bandwidth recall",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "offloading-hierarchy",
      "name": "Offloading Hierarchy",
      "description": "Concept for Offloading hierarchy",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "nvlink-hardware-specs",
      "name": "Nvlink Hardware Specs",
      "description": "Concept for NVLink hardware specs",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "infiniband-ndr-specs",
      "name": "Infiniband Ndr Specs",
      "description": "Concept for InfiniBand NDR specs",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "nvlink-vs-pcie",
      "name": "Nvlink Vs Pcie",
      "description": "Concept for NVLink vs PCIe",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "nvswitch-concept",
      "name": "Nvswitch Concept",
      "description": "Concept for NVSwitch concept",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "data-parallelism-definition",
      "name": "Data Parallelism Definition",
      "description": "Concept for Data Parallelism definition",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "pipeline-parallelism-definition",
      "name": "Pipeline Parallelism Definition",
      "description": "Concept for Pipeline Parallelism definition",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "expert-parallelism-moe",
      "name": "Expert Parallelism (Moe)",
      "description": "Concept for Expert Parallelism (MoE)",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "tpot-metric-definition",
      "name": "Tpot Metric Definition",
      "description": "Concept for TPOT metric definition",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "prefill-vs-decode-phases",
      "name": "Prefill Vs Decode Phases",
      "description": "Concept for Prefill vs Decode phases",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "throughput-vs-latency-tradeoff",
      "name": "Throughput Vs Latency Tradeoff",
      "description": "Concept for Throughput vs Latency tradeoff",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "pagedattention-concept",
      "name": "Pagedattention Concept",
      "description": "Concept for PagedAttention concept",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "kv-cache-size-calculation",
      "name": "Kv Cache Size Calculation",
      "description": "Concept for KV cache size calculation",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "grouped-query-attention-gqa",
      "name": "Grouped Query Attention (Gqa)",
      "description": "Concept for Grouped Query Attention (GQA)",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "arithmetic-intensity-formula",
      "name": "Arithmetic Intensity Formula",
      "description": "Concept for Arithmetic Intensity formula",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "memory-bound-vs-compute-bound",
      "name": "Memory Bound Vs Compute Bound",
      "description": "Concept for Memory bound vs Compute bound",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "ridge-point-calculation",
      "name": "Ridge Point Calculation",
      "description": "Concept for Ridge Point calculation",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "t4-roofline-analysis",
      "name": "T4 Roofline Analysis",
      "description": "Concept for T4 Roofline analysis",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "a100-peak-specifications",
      "name": "A100 Peak Specifications",
      "description": "Concept for A100 Peak specifications",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "operator-fusion-definition",
      "name": "Operator Fusion Definition",
      "description": "Concept for Operator Fusion definition",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "jit-vs-aot-compilation",
      "name": "Jit Vs Aot Compilation",
      "description": "Concept for JIT vs AOT compilation",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "xla-definition",
      "name": "Xla Definition",
      "description": "Concept for XLA definition",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "auto-tuning-definition",
      "name": "Auto Tuning Definition",
      "description": "Concept for Auto tuning definition",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "heartbeat-node-failure-detection",
      "name": "Heartbeat Node Failure Detection",
      "description": "Concept for Heartbeat node failure detection",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "elastic-training-definition",
      "name": "Elastic Training Definition",
      "description": "Concept for Elastic Training definition",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "pipeline-bubble-concept",
      "name": "Pipeline Bubble Concept",
      "description": "Concept for Pipeline Bubble concept",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "quantization-calibration",
      "name": "Quantization Calibration",
      "description": "Concept for quantization calibration",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "operator-fallback",
      "name": "Operator Fallback",
      "description": "Concept for operator fallback",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "batch-norm-folding",
      "name": "Batch Norm Folding",
      "description": "Concept for batch norm folding",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "sub-byte-overhead",
      "name": "Sub Byte Overhead",
      "description": "Concept for sub byte overhead",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "accumulator-overflow",
      "name": "Accumulator Overflow",
      "description": "Concept for accumulator overflow",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "flash-wait-states",
      "name": "Flash Wait States",
      "description": "Concept for flash wait states",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "unstructured-sparsity",
      "name": "Unstructured Sparsity",
      "description": "Concept for unstructured sparsity",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "pcie-bottleneck",
      "name": "Pcie Bottleneck",
      "description": "Concept for pcie bottleneck",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "peak-activation-memory",
      "name": "Peak Activation Memory",
      "description": "Concept for peak activation memory",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "sram-bank-conflict",
      "name": "Sram Bank Conflict",
      "description": "Concept for sram bank conflict",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "cache-thrashing",
      "name": "Cache Thrashing",
      "description": "Concept for cache thrashing",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "flash-power-draw",
      "name": "Flash Power Draw",
      "description": "Concept for flash power draw",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "heap-fragmentation",
      "name": "Heap Fragmentation",
      "description": "Concept for heap fragmentation",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "brownout-reset",
      "name": "Brownout Reset",
      "description": "Concept for brownout reset",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "thermal-throttling",
      "name": "Thermal Throttling",
      "description": "Concept for thermal throttling",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "sram-leakage",
      "name": "Sram Leakage",
      "description": "Concept for sram leakage",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "auto-gain-control",
      "name": "Auto Gain Control",
      "description": "Concept for auto gain control",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "sample-rate-drift",
      "name": "Sample Rate Drift",
      "description": "Concept for sample rate drift",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "blocking-io",
      "name": "Blocking Io",
      "description": "Concept for blocking io",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "ota-partitioning",
      "name": "Ota Partitioning",
      "description": "Concept for ota partitioning",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "watchdog-timeout",
      "name": "Watchdog Timeout",
      "description": "Concept for watchdog timeout",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "sensor-degradation",
      "name": "Sensor Degradation",
      "description": "Concept for sensor degradation",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "interpreter-overhead",
      "name": "Interpreter Overhead",
      "description": "Concept for interpreter overhead",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "endianness-mismatch",
      "name": "Endianness Mismatch",
      "description": "Concept for endianness mismatch",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "buffer-overrun",
      "name": "Buffer Overrun",
      "description": "Concept for buffer overrun",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "control-loop-jitter",
      "name": "Control Loop Jitter",
      "description": "Concept for control loop jitter",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "loop-unrolling-bloat",
      "name": "Loop Unrolling Bloat",
      "description": "Concept for loop unrolling bloat",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "unaligned-memory-access",
      "name": "Unaligned Memory Access",
      "description": "Concept for unaligned memory access",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    },
    {
      "id": "dead-code-elimination",
      "name": "Dead Code Elimination",
      "description": "Concept for dead code elimination",
      "prerequisites": [],
      "tracks": [
        "cloud",
        "edge",
        "mobile",
        "tinyml"
      ],
      "source_chapters": [],
      "source_domains": [],
      "question_count": 0
    }
  ],
  "edges": [
    {
      "source": "allreduce",
      "target": "3d-parallelism",
      "type": "prerequisite"
    },
    {
      "source": "gradient-accumulation",
      "target": "3d-parallelism",
      "type": "prerequisite"
    },
    {
      "source": "mini-batch-processing",
      "target": "3d-parallelism",
      "type": "prerequisite"
    },
    {
      "source": "pipeline-parallelism",
      "target": "3d-parallelism",
      "type": "prerequisite"
    },
    {
      "source": "tensor-parallelism",
      "target": "3d-parallelism",
      "type": "prerequisite"
    },
    {
      "source": "staged-system-pipeline",
      "target": "accelerator-bubbles",
      "type": "prerequisite"
    },
    {
      "source": "activation-memory",
      "target": "activation-checkpointing",
      "type": "prerequisite"
    },
    {
      "source": "compute-bound-workload",
      "target": "activation-checkpointing",
      "type": "prerequisite"
    },
    {
      "source": "memory-bound-workload",
      "target": "activation-checkpointing",
      "type": "prerequisite"
    },
    {
      "source": "reverse-mode-differentiation",
      "target": "activation-checkpointing",
      "type": "prerequisite"
    },
    {
      "source": "stateless-serving",
      "target": "active-active-replication",
      "type": "prerequisite"
    },
    {
      "source": "forward-error-correction",
      "target": "active-optical-cable",
      "type": "prerequisite"
    },
    {
      "source": "elastic-training",
      "target": "adaptive-batch-size-scaling",
      "type": "prerequisite"
    },
    {
      "source": "ecmp",
      "target": "adaptive-routing",
      "type": "prerequisite"
    },
    {
      "source": "adversarial-attacks",
      "target": "adversarial-training",
      "type": "prerequisite"
    },
    {
      "source": "static-computation-graphs",
      "target": "ahead-of-time-optimization",
      "type": "prerequisite"
    },
    {
      "source": "gradient-synchronization",
      "target": "allreduce",
      "type": "prerequisite"
    },
    {
      "source": "bisection-bandwidth",
      "target": "alltoall-communication",
      "type": "prerequisite"
    },
    {
      "source": "embedding-table-sharding",
      "target": "alltoall-communication",
      "type": "prerequisite"
    },
    {
      "source": "latency",
      "target": "alpha-beta-performance-model",
      "type": "prerequisite"
    },
    {
      "source": "computational-intensity",
      "target": "arithmetic-intensity",
      "type": "prerequisite"
    },
    {
      "source": "accelerator-bubbles",
      "target": "asynchronous-prefetching",
      "type": "prerequisite"
    },
    {
      "source": "automatic-differentiation",
      "target": "autograd-tape",
      "type": "prerequisite"
    },
    {
      "source": "eager-execution",
      "target": "autograd-tape",
      "type": "prerequisite"
    },
    {
      "source": "computational-graph",
      "target": "automatic-differentiation",
      "type": "prerequisite"
    },
    {
      "source": "bandwidth-bound-workload",
      "target": "autoregressive-kv-cache",
      "type": "prerequisite"
    },
    {
      "source": "data-poisoning",
      "target": "backdoor-attacks",
      "type": "prerequisite"
    },
    {
      "source": "resource-fragmentation",
      "target": "backfill-scheduling",
      "type": "prerequisite"
    },
    {
      "source": "arithmetic-intensity",
      "target": "bandwidth-bound-workload",
      "type": "prerequisite"
    },
    {
      "source": "latency",
      "target": "bandwidth-distance-product",
      "type": "prerequisite"
    },
    {
      "source": "arithmetic-intensity",
      "target": "bandwidth-hog",
      "type": "prerequisite"
    },
    {
      "source": "memory-bound-workload",
      "target": "bandwidth-hog",
      "type": "prerequisite"
    },
    {
      "source": "activation-memory-footprint",
      "target": "batch-processing-trade-offs",
      "type": "prerequisite"
    },
    {
      "source": "dynamic-batching",
      "target": "batching-window",
      "type": "prerequisite"
    },
    {
      "source": "model-benchmarking",
      "target": "benchmark-contamination",
      "type": "prerequisite"
    },
    {
      "source": "fleet-orchestration",
      "target": "bin-packing-heuristics",
      "type": "prerequisite"
    },
    {
      "source": "model-quantization",
      "target": "binary-neural-networks",
      "type": "prerequisite"
    },
    {
      "source": "precision-engineering",
      "target": "block-wise-quantization",
      "type": "prerequisite"
    },
    {
      "source": "communication-computation-overlap",
      "target": "bucket-fusion",
      "type": "prerequisite"
    },
    {
      "source": "critical-message-size",
      "target": "bucket-fusion",
      "type": "prerequisite"
    },
    {
      "source": "fault-tolerance",
      "target": "byzantine-failures",
      "type": "prerequisite"
    },
    {
      "source": "federated-learning",
      "target": "byzantine-fault-tolerance",
      "type": "prerequisite"
    },
    {
      "source": "cost-per-inference",
      "target": "capacity-planning",
      "type": "prerequisite"
    },
    {
      "source": "service-level-objective",
      "target": "capacity-planning",
      "type": "prerequisite"
    },
    {
      "source": "operational-carbon",
      "target": "carbon-aware-scheduling",
      "type": "prerequisite"
    },
    {
      "source": "on-device-learning",
      "target": "catastrophic-forgetting",
      "type": "prerequisite"
    },
    {
      "source": "checkpointing",
      "target": "centralized-checkpointing",
      "type": "prerequisite"
    },
    {
      "source": "adversarial-attacks",
      "target": "certified-robustness",
      "type": "prerequisite"
    },
    {
      "source": "parallel-file-system",
      "target": "checkpoint-storms",
      "type": "prerequisite"
    },
    {
      "source": "sequential-throughput",
      "target": "checkpoint-storms",
      "type": "prerequisite"
    },
    {
      "source": "fault-tolerance",
      "target": "checkpointing",
      "type": "prerequisite"
    },
    {
      "source": "reticle-limit",
      "target": "chiplet-architecture",
      "type": "prerequisite"
    },
    {
      "source": "data-poisoning",
      "target": "clean-label-poisoning",
      "type": "prerequisite"
    },
    {
      "source": "checkpointing",
      "target": "cold-restart",
      "type": "prerequisite"
    },
    {
      "source": "communication-computation-ratio",
      "target": "communication-bound",
      "type": "prerequisite"
    },
    {
      "source": "cuda-streams",
      "target": "communication-computation-overlap",
      "type": "prerequisite"
    },
    {
      "source": "eager-execution",
      "target": "compilation-continuum-principle",
      "type": "prerequisite"
    },
    {
      "source": "just-in-time-compilation",
      "target": "compilation-continuum-principle",
      "type": "prerequisite"
    },
    {
      "source": "processing-elements",
      "target": "computation-placement",
      "type": "prerequisite"
    },
    {
      "source": "training-cost-asymmetry",
      "target": "computational-intensity",
      "type": "prerequisite"
    },
    {
      "source": "communication-computation-ratio",
      "target": "compute-bound",
      "type": "prerequisite"
    },
    {
      "source": "memory-wall",
      "target": "compute-bound-workload",
      "type": "prerequisite"
    },
    {
      "source": "covariate-shift",
      "target": "concept-drift",
      "type": "prerequisite"
    },
    {
      "source": "distribution-shift",
      "target": "concept-drift",
      "type": "prerequisite"
    },
    {
      "source": "silent-model-failures",
      "target": "concept-drift",
      "type": "prerequisite"
    },
    {
      "source": "secure-model-design",
      "target": "confidence-calibration",
      "type": "prerequisite"
    },
    {
      "source": "priority-flow-control",
      "target": "congestion-spreading",
      "type": "prerequisite"
    },
    {
      "source": "elastic-training",
      "target": "constant-batch-size-scaling",
      "type": "prerequisite"
    },
    {
      "source": "stage-interface-contracts",
      "target": "constraint-propagation",
      "type": "prerequisite"
    },
    {
      "source": "model-flops-utilization",
      "target": "continuous-batching",
      "type": "prerequisite"
    },
    {
      "source": "static-batching",
      "target": "continuous-batching",
      "type": "prerequisite"
    },
    {
      "source": "self-supervised-learning",
      "target": "contrastive-learning",
      "type": "prerequisite"
    },
    {
      "source": "boundary-erosion",
      "target": "correction-cascades",
      "type": "prerequisite"
    },
    {
      "source": "failure-rate",
      "target": "correlated-failures",
      "type": "prerequisite"
    },
    {
      "source": "distribution-shift",
      "target": "covariate-shift",
      "type": "prerequisite"
    },
    {
      "source": "kl-divergence",
      "target": "covariate-shift",
      "type": "prerequisite"
    },
    {
      "source": "latency-jitter",
      "target": "cpu-affinity",
      "type": "prerequisite"
    },
    {
      "source": "alpha-beta-performance-model",
      "target": "critical-message-size",
      "type": "prerequisite"
    },
    {
      "source": "kernel-launch-overhead",
      "target": "cuda-graphs",
      "type": "prerequisite"
    },
    {
      "source": "accelerator-spectrum",
      "target": "custom-asic",
      "type": "prerequisite"
    },
    {
      "source": "data-benchmarking",
      "target": "data-centric-ai",
      "type": "prerequisite"
    },
    {
      "source": "data-gravity",
      "target": "data-locality-invariant",
      "type": "prerequisite"
    },
    {
      "source": "light-barrier",
      "target": "data-locality-invariant",
      "type": "prerequisite"
    },
    {
      "source": "memory-wall",
      "target": "data-movement-energy-cost",
      "type": "prerequisite"
    },
    {
      "source": "i-o-wall",
      "target": "data-pipeline-equation",
      "type": "prerequisite"
    },
    {
      "source": "sequential-throughput",
      "target": "data-sharding",
      "type": "prerequisite"
    },
    {
      "source": "data-pipeline-equation",
      "target": "data-stall-ratio",
      "type": "prerequisite"
    },
    {
      "source": "computation-placement",
      "target": "dataflow-optimization",
      "type": "prerequisite"
    },
    {
      "source": "memory-allocation",
      "target": "dataflow-optimization",
      "type": "prerequisite"
    },
    {
      "source": "feeding-tax",
      "target": "dataloader-choke-point",
      "type": "prerequisite"
    },
    {
      "source": "roce",
      "target": "dcqcn",
      "type": "prerequisite"
    },
    {
      "source": "schema-validation",
      "target": "dead-letter-queues",
      "type": "prerequisite"
    },
    {
      "source": "fleet-orchestration",
      "target": "declarative-scheduling",
      "type": "prerequisite"
    },
    {
      "source": "system-entropy",
      "target": "degradation-equation",
      "type": "prerequisite"
    },
    {
      "source": "disaggregated-evaluation",
      "target": "demographic-parity",
      "type": "prerequisite"
    },
    {
      "source": "model-registry",
      "target": "dependency-aware-deployment",
      "type": "prerequisite"
    },
    {
      "source": "arithmetic-intensity",
      "target": "depthwise-separable-convolution",
      "type": "prerequisite"
    },
    {
      "source": "federated-learning",
      "target": "differential-privacy",
      "type": "prerequisite"
    },
    {
      "source": "time-per-output-token",
      "target": "disaggregated-serving",
      "type": "prerequisite"
    },
    {
      "source": "time-to-first-token",
      "target": "disaggregated-serving",
      "type": "prerequisite"
    },
    {
      "source": "demographic-parity",
      "target": "disparate-impact-ratio",
      "type": "prerequisite"
    },
    {
      "source": "dispatch-tax",
      "target": "dispatch-overhead-law",
      "type": "prerequisite"
    },
    {
      "source": "eager-execution",
      "target": "dispatch-tax",
      "type": "prerequisite"
    },
    {
      "source": "checkpointing",
      "target": "distributed-checkpointing",
      "type": "prerequisite"
    },
    {
      "source": "data-benchmarking",
      "target": "distribution-shift",
      "type": "prerequisite"
    },
    {
      "source": "stationarity-assumption",
      "target": "distribution-shift",
      "type": "prerequisite"
    },
    {
      "source": "bisection-bandwidth",
      "target": "dragonfly-topology",
      "type": "prerequisite"
    },
    {
      "source": "activation-sparsity",
      "target": "dying-relu-problem",
      "type": "prerequisite"
    },
    {
      "source": "dense-matrix-multiplication",
      "target": "dynamic-attention-routing",
      "type": "prerequisite"
    },
    {
      "source": "static-batching",
      "target": "dynamic-batching",
      "type": "prerequisite"
    },
    {
      "source": "inference-server",
      "target": "dynamic-batching",
      "type": "prerequisite"
    },
    {
      "source": "latency",
      "target": "dynamic-batching",
      "type": "prerequisite"
    },
    {
      "source": "throughput",
      "target": "dynamic-batching",
      "type": "prerequisite"
    },
    {
      "source": "data-benchmarking",
      "target": "dynamic-benchmarking",
      "type": "prerequisite"
    },
    {
      "source": "inference-pipeline",
      "target": "dynamic-inference",
      "type": "prerequisite"
    },
    {
      "source": "latency",
      "target": "dynamic-inference",
      "type": "prerequisite"
    },
    {
      "source": "cmos-power-equation",
      "target": "dynamic-power",
      "type": "prerequisite"
    },
    {
      "source": "thermal-design-power",
      "target": "dynamic-voltage-frequency-scaling",
      "type": "prerequisite"
    },
    {
      "source": "sutva-violations",
      "target": "ego-exclusion-design",
      "type": "prerequisite"
    },
    {
      "source": "checkpointing",
      "target": "elastic-training",
      "type": "prerequisite"
    },
    {
      "source": "fleet-orchestration",
      "target": "elastic-training",
      "type": "prerequisite"
    },
    {
      "source": "catastrophic-forgetting",
      "target": "elastic-weight-consolidation",
      "type": "prerequisite"
    },
    {
      "source": "batch-ingestion",
      "target": "elt-processing",
      "type": "prerequisite"
    },
    {
      "source": "model-parallelism",
      "target": "embedding-table-sharding",
      "type": "prerequisite"
    },
    {
      "source": "sparse-embedding-lookups",
      "target": "embedding-table-sharding",
      "type": "prerequisite"
    },
    {
      "source": "system-benchmarking",
      "target": "energy-benchmarking",
      "type": "prerequisite"
    },
    {
      "source": "data-gravity",
      "target": "energy-movement-invariant",
      "type": "prerequisite"
    },
    {
      "source": "arithmetic-intensity",
      "target": "energy-roofline-model",
      "type": "prerequisite"
    },
    {
      "source": "energy-per-byte",
      "target": "energy-roofline-model",
      "type": "prerequisite"
    },
    {
      "source": "energy-per-operation",
      "target": "energy-roofline-model",
      "type": "prerequisite"
    },
    {
      "source": "equalized-odds",
      "target": "equality-of-opportunity",
      "type": "prerequisite"
    },
    {
      "source": "object-storage-streaming",
      "target": "erasure-coding",
      "type": "prerequisite"
    },
    {
      "source": "transient-failures",
      "target": "error-correcting-codes",
      "type": "prerequisite"
    },
    {
      "source": "gradient-compression",
      "target": "error-feedback",
      "type": "prerequisite"
    },
    {
      "source": "fault-injection",
      "target": "error-model",
      "type": "prerequisite"
    },
    {
      "source": "pinned-memory",
      "target": "execution-streams",
      "type": "prerequisite"
    },
    {
      "source": "model-benchmarking",
      "target": "expected-calibration-error",
      "type": "prerequisite"
    },
    {
      "source": "catastrophic-forgetting",
      "target": "experience-replay",
      "type": "prerequisite"
    },
    {
      "source": "pre-allocated-memory-management",
      "target": "external-fragmentation",
      "type": "prerequisite"
    },
    {
      "source": "fault-tolerance",
      "target": "fail-stop-failures",
      "type": "prerequisite"
    },
    {
      "source": "correlated-failures",
      "target": "failure-domains",
      "type": "prerequisite"
    },
    {
      "source": "fault-tolerance",
      "target": "failure-rate",
      "type": "prerequisite"
    },
    {
      "source": "fleet-orchestration",
      "target": "fair-share-scheduling",
      "type": "prerequisite"
    },
    {
      "source": "demographic-parity",
      "target": "fairness-impossibility-law",
      "type": "prerequisite"
    },
    {
      "source": "equalized-odds",
      "target": "fairness-impossibility-law",
      "type": "prerequisite"
    },
    {
      "source": "bisection-bandwidth",
      "target": "fat-tree-topology",
      "type": "prerequisite"
    },
    {
      "source": "fault-tolerance",
      "target": "fault-injection",
      "type": "prerequisite"
    },
    {
      "source": "physical-tampering",
      "target": "fault-injection-attack",
      "type": "prerequisite"
    },
    {
      "source": "fault-injection",
      "target": "fault-model",
      "type": "prerequisite"
    },
    {
      "source": "training-serving-skew",
      "target": "feature-store",
      "type": "prerequisite"
    },
    {
      "source": "federated-learning",
      "target": "federated-analytics",
      "type": "prerequisite"
    },
    {
      "source": "secure-aggregation",
      "target": "federated-analytics",
      "type": "prerequisite"
    },
    {
      "source": "on-device-learning",
      "target": "federated-learning",
      "type": "prerequisite"
    },
    {
      "source": "data-gravity",
      "target": "feeding-tax",
      "type": "prerequisite"
    },
    {
      "source": "model-benchmarking",
      "target": "time-to-first-token",
      "type": "prerequisite"
    },
    {
      "source": "memory-bound-workload",
      "target": "flashattention",
      "type": "prerequisite"
    },
    {
      "source": "online-softmax",
      "target": "flashattention",
      "type": "prerequisite"
    },
    {
      "source": "operator-fusion",
      "target": "flashattention",
      "type": "prerequisite"
    },
    {
      "source": "tiling-algorithms",
      "target": "flashattention",
      "type": "prerequisite"
    },
    {
      "source": "pam4",
      "target": "forward-error-correction",
      "type": "prerequisite"
    },
    {
      "source": "automatic-differentiation",
      "target": "forward-mode-differentiation",
      "type": "prerequisite"
    },
    {
      "source": "fleet-orchestration",
      "target": "gang-scheduling",
      "type": "prerequisite"
    },
    {
      "source": "accelerator-spectrum",
      "target": "generality-tax",
      "type": "prerequisite"
    },
    {
      "source": "high-bandwidth-memory",
      "target": "gpu-direct-storage",
      "type": "prerequisite"
    },
    {
      "source": "local-nvme-warm-cache",
      "target": "gpu-direct-storage",
      "type": "prerequisite"
    },
    {
      "source": "preprocessing-bottleneck",
      "target": "gpu-preprocessing",
      "type": "prerequisite"
    },
    {
      "source": "remote-direct-memory-access",
      "target": "gpudirect-rdma",
      "type": "prerequisite"
    },
    {
      "source": "activation-memory",
      "target": "gradient-accumulation",
      "type": "prerequisite"
    },
    {
      "source": "mini-batch-processing",
      "target": "gradient-accumulation",
      "type": "prerequisite"
    },
    {
      "source": "gradient-synchronization",
      "target": "gradient-compression",
      "type": "prerequisite"
    },
    {
      "source": "federated-learning",
      "target": "gradient-inversion-attack",
      "type": "prerequisite"
    },
    {
      "source": "federated-learning",
      "target": "gradient-quantization",
      "type": "prerequisite"
    },
    {
      "source": "just-in-time-compilation",
      "target": "graph-breaks",
      "type": "prerequisite"
    },
    {
      "source": "sutva-violations",
      "target": "graph-cluster-randomization",
      "type": "prerequisite"
    },
    {
      "source": "operator-fusion",
      "target": "graph-compilation",
      "type": "prerequisite"
    },
    {
      "source": "just-in-time-compilation",
      "target": "graph-scripting",
      "type": "prerequisite"
    },
    {
      "source": "just-in-time-compilation",
      "target": "graph-tracing",
      "type": "prerequisite"
    },
    {
      "source": "spurious-correlations",
      "target": "group-distributionally-robust-optimization",
      "type": "prerequisite"
    },
    {
      "source": "kv-cache",
      "target": "grouped-query-attention",
      "type": "prerequisite"
    },
    {
      "source": "zero-copy-serialization",
      "target": "grpc-protocol",
      "type": "prerequisite"
    },
    {
      "source": "tensor-cores",
      "target": "hardware-precision-scaling",
      "type": "prerequisite"
    },
    {
      "source": "fault-tolerance",
      "target": "hardware-redundancy",
      "type": "prerequisite"
    },
    {
      "source": "physical-tampering",
      "target": "hardware-trojans",
      "type": "prerequisite"
    },
    {
      "source": "tail-latency",
      "target": "hedged-requests",
      "type": "prerequisite"
    },
    {
      "source": "allreduce",
      "target": "hierarchical-allreduce",
      "type": "prerequisite"
    },
    {
      "source": "ring-allreduce",
      "target": "hierarchical-allreduce",
      "type": "prerequisite"
    },
    {
      "source": "tinyml",
      "target": "hierarchical-processing",
      "type": "prerequisite"
    },
    {
      "source": "gpu-memory-hierarchy",
      "target": "high-bandwidth-memory",
      "type": "prerequisite"
    },
    {
      "source": "memory-wall",
      "target": "high-bandwidth-memory",
      "type": "prerequisite"
    },
    {
      "source": "ml-storage-hierarchy",
      "target": "high-bandwidth-memory",
      "type": "prerequisite"
    },
    {
      "source": "gang-scheduling",
      "target": "hold-and-wait-deadlock",
      "type": "prerequisite"
    },
    {
      "source": "high-bandwidth-memory",
      "target": "host-dram-staging",
      "type": "prerequisite"
    },
    {
      "source": "declarative-scheduling",
      "target": "host-networking-bypass",
      "type": "prerequisite"
    },
    {
      "source": "tail-latency",
      "target": "i-o-jitter",
      "type": "prerequisite"
    },
    {
      "source": "sequential-throughput",
      "target": "i-o-wall",
      "type": "prerequisite"
    },
    {
      "source": "dense-matrix-multiplication",
      "target": "im2col-transformation",
      "type": "prerequisite"
    },
    {
      "source": "matrix-operations",
      "target": "im2col-transformation",
      "type": "prerequisite"
    },
    {
      "source": "fleet-orchestration",
      "target": "imperative-scheduling",
      "type": "prerequisite"
    },
    {
      "source": "bulk-synchronous-parallel",
      "target": "incast",
      "type": "prerequisite"
    },
    {
      "source": "latency",
      "target": "inference-server",
      "type": "prerequisite"
    },
    {
      "source": "throughput",
      "target": "inference-server",
      "type": "prerequisite"
    },
    {
      "source": "remote-direct-memory-access",
      "target": "infiniband",
      "type": "prerequisite"
    },
    {
      "source": "dataflow-optimization",
      "target": "input-stationary",
      "type": "prerequisite"
    },
    {
      "source": "memory-wall",
      "target": "integration-bottleneck",
      "type": "prerequisite"
    },
    {
      "source": "latency-jitter",
      "target": "interrupt-shielding",
      "type": "prerequisite"
    },
    {
      "source": "demographic-parity",
      "target": "intersectional-fairness",
      "type": "prerequisite"
    },
    {
      "source": "arithmetic-intensity",
      "target": "io-aware-algorithm-design",
      "type": "prerequisite"
    },
    {
      "source": "memory-bound-workload",
      "target": "io-aware-algorithm-design",
      "type": "prerequisite"
    },
    {
      "source": "memory-wall",
      "target": "iron-law-of-ml-systems",
      "type": "prerequisite"
    },
    {
      "source": "power-wall",
      "target": "iron-law-of-ml-systems",
      "type": "prerequisite"
    },
    {
      "source": "accelerator-bubbles",
      "target": "iron-law-of-training",
      "type": "prerequisite"
    },
    {
      "source": "computational-intensity",
      "target": "iron-law-of-training",
      "type": "prerequisite"
    },
    {
      "source": "constraint-propagation",
      "target": "iteration-tax",
      "type": "prerequisite"
    },
    {
      "source": "kl-divergence",
      "target": "jensen-shannon-divergence",
      "type": "prerequisite"
    },
    {
      "source": "computational-graph",
      "target": "just-in-time-compilation",
      "type": "prerequisite"
    },
    {
      "source": "eager-execution",
      "target": "just-in-time-compilation",
      "type": "prerequisite"
    },
    {
      "source": "computational-graph",
      "target": "kernel-fusion",
      "type": "prerequisite"
    },
    {
      "source": "memory-bound-workload",
      "target": "kernel-fusion",
      "type": "prerequisite"
    },
    {
      "source": "memory-boundedness",
      "target": "kernel-fusion",
      "type": "prerequisite"
    },
    {
      "source": "distribution-shift",
      "target": "kl-divergence",
      "type": "prerequisite"
    },
    {
      "source": "hardware-memory-hierarchy",
      "target": "kv-cache-bottleneck",
      "type": "prerequisite"
    },
    {
      "source": "stateful-serving",
      "target": "kv-cache-checkpointing",
      "type": "prerequisite"
    },
    {
      "source": "kv-cache",
      "target": "kv-cache-quantization",
      "type": "prerequisite"
    },
    {
      "source": "concept-drift",
      "target": "label-quality-drift",
      "type": "prerequisite"
    },
    {
      "source": "covariate-shift",
      "target": "label-shift",
      "type": "prerequisite"
    },
    {
      "source": "distribution-shift",
      "target": "label-shift",
      "type": "prerequisite"
    },
    {
      "source": "inference-pipeline",
      "target": "latency-budget",
      "type": "prerequisite"
    },
    {
      "source": "service-level-objective",
      "target": "latency-budget",
      "type": "prerequisite"
    },
    {
      "source": "latency",
      "target": "latency-jitter",
      "type": "prerequisite"
    },
    {
      "source": "tail-latency",
      "target": "latency-jitter",
      "type": "prerequisite"
    },
    {
      "source": "embodied-carbon",
      "target": "lifecycle-carbon-accounting",
      "type": "prerequisite"
    },
    {
      "source": "operational-carbon",
      "target": "lifecycle-carbon-accounting",
      "type": "prerequisite"
    },
    {
      "source": "data-sharding",
      "target": "local-nvme-warm-cache",
      "type": "prerequisite"
    },
    {
      "source": "local-nvme-warm-cache",
      "target": "locality-aware-scheduling",
      "type": "prerequisite"
    },
    {
      "source": "on-device-learning",
      "target": "low-rank-adapters",
      "type": "prerequisite"
    },
    {
      "source": "multiply-accumulate-operations",
      "target": "matrix-multiplication-workloads",
      "type": "prerequisite"
    },
    {
      "source": "distribution-shift",
      "target": "maximum-mean-discrepancy",
      "type": "prerequisite"
    },
    {
      "source": "differential-privacy",
      "target": "membership-inference-attack",
      "type": "prerequisite"
    },
    {
      "source": "model-memorization",
      "target": "membership-inference-attack",
      "type": "prerequisite"
    },
    {
      "source": "memory-wall",
      "target": "memory-bound-workload",
      "type": "prerequisite"
    },
    {
      "source": "memory-wall",
      "target": "memory-boundedness",
      "type": "prerequisite"
    },
    {
      "source": "latency-jitter",
      "target": "memory-locking",
      "type": "prerequisite"
    },
    {
      "source": "activation-memory",
      "target": "memory-pressure",
      "type": "prerequisite"
    },
    {
      "source": "optimizer-memory-overhead",
      "target": "memory-pressure",
      "type": "prerequisite"
    },
    {
      "source": "gpu-memory-hierarchy",
      "target": "memory-wall",
      "type": "prerequisite"
    },
    {
      "source": "compute-bound-workload",
      "target": "mixed-precision-training",
      "type": "prerequisite"
    },
    {
      "source": "memory-pressure",
      "target": "mixed-precision-training",
      "type": "prerequisite"
    },
    {
      "source": "reverse-mode-differentiation",
      "target": "mixed-precision-training",
      "type": "prerequisite"
    },
    {
      "source": "i-o-wall",
      "target": "ml-storage-hierarchy",
      "type": "prerequisite"
    },
    {
      "source": "disaggregated-evaluation",
      "target": "model-cards",
      "type": "prerequisite"
    },
    {
      "source": "high-bandwidth-memory",
      "target": "model-cold-start-latency",
      "type": "prerequisite"
    },
    {
      "source": "sequential-throughput",
      "target": "model-cold-start-latency",
      "type": "prerequisite"
    },
    {
      "source": "distribution-shift",
      "target": "model-ensembles",
      "type": "prerequisite"
    },
    {
      "source": "hardware-flops-utilization",
      "target": "model-flops-utilization",
      "type": "prerequisite"
    },
    {
      "source": "roofline-model",
      "target": "model-flops-utilization",
      "type": "prerequisite"
    },
    {
      "source": "memory-pressure",
      "target": "model-parallelism",
      "type": "prerequisite"
    },
    {
      "source": "dynamic-power",
      "target": "model-quantization",
      "type": "prerequisite"
    },
    {
      "source": "energy-per-byte",
      "target": "model-quantization",
      "type": "prerequisite"
    },
    {
      "source": "memory-bound-workload",
      "target": "model-quantization",
      "type": "prerequisite"
    },
    {
      "source": "memory-boundedness",
      "target": "model-quantization",
      "type": "prerequisite"
    },
    {
      "source": "model-format-optimization",
      "target": "model-quantization",
      "type": "prerequisite"
    },
    {
      "source": "secure-model-design",
      "target": "model-watermarking",
      "type": "prerequisite"
    },
    {
      "source": "resource-fragmentation",
      "target": "multi-instance-gpu-mig",
      "type": "prerequisite"
    },
    {
      "source": "canary-deployment",
      "target": "multi-region-deployment",
      "type": "prerequisite"
    },
    {
      "source": "prefetch-buffer-depth",
      "target": "multi-worker-data-loading",
      "type": "prerequisite"
    },
    {
      "source": "intersectional-fairness",
      "target": "multicalibration",
      "type": "prerequisite"
    },
    {
      "source": "federated-learning",
      "target": "non-iid-data-distributions",
      "type": "prerequisite"
    },
    {
      "source": "host-dram-staging",
      "target": "numa-aware-allocation",
      "type": "prerequisite"
    },
    {
      "source": "cpu-affinity",
      "target": "numa-aware-inference",
      "type": "prerequisite"
    },
    {
      "source": "data-sharding",
      "target": "object-storage-streaming",
      "type": "prerequisite"
    },
    {
      "source": "concept-drift",
      "target": "online-learning",
      "type": "prerequisite"
    },
    {
      "source": "dataflow-optimization",
      "target": "operator-fusion",
      "type": "prerequisite"
    },
    {
      "source": "gpu-memory-hierarchy",
      "target": "operator-fusion",
      "type": "prerequisite"
    },
    {
      "source": "memory-wall",
      "target": "operator-fusion",
      "type": "prerequisite"
    },
    {
      "source": "on-device-learning",
      "target": "opportunistic-training-scheduling",
      "type": "prerequisite"
    },
    {
      "source": "dataflow-optimization",
      "target": "output-stationary",
      "type": "prerequisite"
    },
    {
      "source": "pre-allocated-memory-management",
      "target": "paged-memory-management",
      "type": "prerequisite"
    },
    {
      "source": "kv-cache",
      "target": "pagedattention-mechanism",
      "type": "prerequisite"
    },
    {
      "source": "paged-memory-management",
      "target": "pagedattention-mechanism",
      "type": "prerequisite"
    },
    {
      "source": "ml-storage-hierarchy",
      "target": "parallel-file-system",
      "type": "prerequisite"
    },
    {
      "source": "model-benchmarking",
      "target": "pareto-frontier",
      "type": "prerequisite"
    },
    {
      "source": "on-device-learning",
      "target": "personalization-layers",
      "type": "prerequisite"
    },
    {
      "source": "tensor-abstraction",
      "target": "pinned-memory",
      "type": "prerequisite"
    },
    {
      "source": "pipeline-parallelism",
      "target": "pipeline-bubble",
      "type": "prerequisite"
    },
    {
      "source": "accelerator-bubbles",
      "target": "pipeline-parallelism",
      "type": "prerequisite"
    },
    {
      "source": "model-parallelism",
      "target": "pipeline-parallelism",
      "type": "prerequisite"
    },
    {
      "source": "distribution-shift",
      "target": "population-stability-index",
      "type": "prerequisite"
    },
    {
      "source": "schema-validation",
      "target": "population-stability-index",
      "type": "prerequisite"
    },
    {
      "source": "system-explainability",
      "target": "post-hoc-explainability",
      "type": "prerequisite"
    },
    {
      "source": "precision-engineering",
      "target": "post-training-quantization",
      "type": "prerequisite"
    },
    {
      "source": "side-channel-attack",
      "target": "power-analysis-attack",
      "type": "prerequisite"
    },
    {
      "source": "dennard-scaling",
      "target": "power-wall",
      "type": "prerequisite"
    },
    {
      "source": "memory-allocation",
      "target": "pre-allocated-memory-management",
      "type": "prerequisite"
    },
    {
      "source": "memory-wall",
      "target": "precision-engineering",
      "type": "prerequisite"
    },
    {
      "source": "host-dram-staging",
      "target": "prefetch-buffer-depth",
      "type": "prerequisite"
    },
    {
      "source": "i-o-jitter",
      "target": "prefetch-buffer-depth",
      "type": "prerequisite"
    },
    {
      "source": "arithmetic-intensity",
      "target": "prefill-decode-decomposition",
      "type": "prerequisite"
    },
    {
      "source": "pagedattention-mechanism",
      "target": "prefix-caching",
      "type": "prerequisite"
    },
    {
      "source": "latency-budget",
      "target": "preprocessing-bottleneck",
      "type": "prerequisite"
    },
    {
      "source": "go-back-n",
      "target": "priority-flow-control",
      "type": "prerequisite"
    },
    {
      "source": "roce",
      "target": "priority-flow-control",
      "type": "prerequisite"
    },
    {
      "source": "hold-and-wait-deadlock",
      "target": "priority-inversion",
      "type": "prerequisite"
    },
    {
      "source": "3d-parallelism",
      "target": "process-group",
      "type": "prerequisite"
    },
    {
      "source": "hierarchical-processing",
      "target": "progressive-deployment",
      "type": "prerequisite"
    },
    {
      "source": "model-quantization",
      "target": "progressive-deployment",
      "type": "prerequisite"
    },
    {
      "source": "dynamic-attention-routing",
      "target": "quadratic-attention-bottleneck",
      "type": "prerequisite"
    },
    {
      "source": "precision-engineering",
      "target": "quantization-aware-training",
      "type": "prerequisite"
    },
    {
      "source": "topology-aware-placement",
      "target": "rail-optimized-scheduling",
      "type": "prerequisite"
    },
    {
      "source": "tensor-parallelism",
      "target": "rail-optimized-topology",
      "type": "prerequisite"
    },
    {
      "source": "cost-per-inference",
      "target": "reactive-scaling",
      "type": "prerequisite"
    },
    {
      "source": "demographic-parity",
      "target": "real-time-fairness-monitoring",
      "type": "prerequisite"
    },
    {
      "source": "bin-packing-heuristics",
      "target": "resource-fragmentation",
      "type": "prerequisite"
    },
    {
      "source": "serialization-overhead",
      "target": "rest-protocol",
      "type": "prerequisite"
    },
    {
      "source": "computational-graph",
      "target": "reverse-linked-graph",
      "type": "prerequisite"
    },
    {
      "source": "reverse-mode-differentiation",
      "target": "reverse-linked-graph",
      "type": "prerequisite"
    },
    {
      "source": "automatic-differentiation",
      "target": "reverse-mode-differentiation",
      "type": "prerequisite"
    },
    {
      "source": "allreduce",
      "target": "ring-allreduce",
      "type": "prerequisite"
    },
    {
      "source": "remote-direct-memory-access",
      "target": "roce",
      "type": "prerequisite"
    },
    {
      "source": "arithmetic-intensity",
      "target": "roofline-model",
      "type": "prerequisite"
    },
    {
      "source": "memory-boundedness",
      "target": "roofline-model",
      "type": "prerequisite"
    },
    {
      "source": "memory-wall",
      "target": "roofline-model",
      "type": "prerequisite"
    },
    {
      "source": "layered-defense",
      "target": "runtime-input-validation",
      "type": "prerequisite"
    },
    {
      "source": "layered-defense",
      "target": "runtime-output-monitoring",
      "type": "prerequisite"
    },
    {
      "source": "system-explainability",
      "target": "saliency-maps",
      "type": "prerequisite"
    },
    {
      "source": "elastic-training",
      "target": "scaling-efficiency-curve",
      "type": "prerequisite"
    },
    {
      "source": "static-batching",
      "target": "scaling-efficiency-curve",
      "type": "prerequisite"
    },
    {
      "source": "federated-learning",
      "target": "secure-aggregation",
      "type": "prerequisite"
    },
    {
      "source": "gradient-inversion-attack",
      "target": "secure-aggregation",
      "type": "prerequisite"
    },
    {
      "source": "layered-defense",
      "target": "secure-model-deployment",
      "type": "prerequisite"
    },
    {
      "source": "layered-defense",
      "target": "secure-model-design",
      "type": "prerequisite"
    },
    {
      "source": "on-device-learning",
      "target": "self-supervised-learning",
      "type": "prerequisite"
    },
    {
      "source": "data-sharding",
      "target": "sequential-streaming-formats",
      "type": "prerequisite"
    },
    {
      "source": "metadata-overhead",
      "target": "sequential-streaming-formats",
      "type": "prerequisite"
    },
    {
      "source": "latency",
      "target": "service-level-objective",
      "type": "prerequisite"
    },
    {
      "source": "tail-latency",
      "target": "service-level-objective",
      "type": "prerequisite"
    },
    {
      "source": "latency",
      "target": "serving-inversion",
      "type": "prerequisite"
    },
    {
      "source": "throughput",
      "target": "serving-inversion",
      "type": "prerequisite"
    },
    {
      "source": "on-device-learning",
      "target": "shadow-model-evaluation",
      "type": "prerequisite"
    },
    {
      "source": "data-sharding",
      "target": "shard-level-shuffling",
      "type": "prerequisite"
    },
    {
      "source": "distributed-checkpointing",
      "target": "sharded-checkpointing",
      "type": "prerequisite"
    },
    {
      "source": "byzantine-failures",
      "target": "silent-data-corruption",
      "type": "prerequisite"
    },
    {
      "source": "transient-failures",
      "target": "silent-data-corruption",
      "type": "prerequisite"
    },
    {
      "source": "system-benchmarking",
      "target": "silent-degradation",
      "type": "prerequisite"
    },
    {
      "source": "high-bandwidth-memory",
      "target": "silicon-interposer",
      "type": "prerequisite"
    },
    {
      "source": "vector-operations",
      "target": "simd-execution",
      "type": "prerequisite"
    },
    {
      "source": "remote-direct-memory-access",
      "target": "single-root-i-o-virtualization",
      "type": "prerequisite"
    },
    {
      "source": "just-in-time-compilation",
      "target": "single-static-assignment",
      "type": "prerequisite"
    },
    {
      "source": "metadata-overhead",
      "target": "small-file-problem",
      "type": "prerequisite"
    },
    {
      "source": "random-iops",
      "target": "small-file-problem",
      "type": "prerequisite"
    },
    {
      "source": "block-wise-quantization",
      "target": "smoothquant",
      "type": "prerequisite"
    },
    {
      "source": "precision-engineering",
      "target": "smoothquant",
      "type": "prerequisite"
    },
    {
      "source": "memory-capacity-bound-workload",
      "target": "sparse-embedding-lookups",
      "type": "prerequisite"
    },
    {
      "source": "federated-learning",
      "target": "sparse-gradient-updates",
      "type": "prerequisite"
    },
    {
      "source": "memory-bound-workload",
      "target": "sparse-scatter",
      "type": "prerequisite"
    },
    {
      "source": "data-poisoning",
      "target": "spectral-signatures",
      "type": "prerequisite"
    },
    {
      "source": "fleet-orchestration",
      "target": "spot-instances",
      "type": "prerequisite"
    },
    {
      "source": "continuous-training-pipelines",
      "target": "staleness-cost-function",
      "type": "prerequisite"
    },
    {
      "source": "federated-learning",
      "target": "state-reconciliation",
      "type": "prerequisite"
    },
    {
      "source": "fault-tolerance",
      "target": "stateful-serving",
      "type": "prerequisite"
    },
    {
      "source": "fault-tolerance",
      "target": "stateless-serving",
      "type": "prerequisite"
    },
    {
      "source": "computational-graph",
      "target": "static-computation-graphs",
      "type": "prerequisite"
    },
    {
      "source": "inference-pipeline",
      "target": "static-inference",
      "type": "prerequisite"
    },
    {
      "source": "cmos-power-equation",
      "target": "static-power",
      "type": "prerequisite"
    },
    {
      "source": "demographic-parity",
      "target": "statistical-parity-difference",
      "type": "prerequisite"
    },
    {
      "source": "multi-region-deployment",
      "target": "sticky-routing",
      "type": "prerequisite"
    },
    {
      "source": "fault-tolerance",
      "target": "straggler-mitigation",
      "type": "prerequisite"
    },
    {
      "source": "federated-learning",
      "target": "straggler-mitigation",
      "type": "prerequisite"
    },
    {
      "source": "dynamic-batching",
      "target": "streaming-inference",
      "type": "prerequisite"
    },
    {
      "source": "checkpointing",
      "target": "synchronous-checkpointing",
      "type": "prerequisite"
    },
    {
      "source": "matrix-operations",
      "target": "systolic-array",
      "type": "prerequisite"
    },
    {
      "source": "latency",
      "target": "tail-latency",
      "type": "prerequisite"
    },
    {
      "source": "bare-metal-serving",
      "target": "tensor-arena",
      "type": "prerequisite"
    },
    {
      "source": "matrix-operations",
      "target": "tensor-cores",
      "type": "prerequisite"
    },
    {
      "source": "inference-server",
      "target": "tensor-format-transformation",
      "type": "prerequisite"
    },
    {
      "source": "model-parallelism",
      "target": "tensor-parallelism",
      "type": "prerequisite"
    },
    {
      "source": "model-sharding",
      "target": "tensor-parallelism",
      "type": "prerequisite"
    },
    {
      "source": "tensor-abstraction",
      "target": "tensor-stride-patterns",
      "type": "prerequisite"
    },
    {
      "source": "matrix-multiplication-workloads",
      "target": "memory-wall",
      "type": "prerequisite"
    },
    {
      "source": "equality-of-opportunity",
      "target": "threshold-adjustment",
      "type": "prerequisite"
    },
    {
      "source": "high-bandwidth-memory",
      "target": "through-silicon-vias",
      "type": "prerequisite"
    },
    {
      "source": "io-aware-algorithm-design",
      "target": "tiling-algorithms",
      "type": "prerequisite"
    },
    {
      "source": "time-to-first-token",
      "target": "time-per-output-token",
      "type": "prerequisite"
    },
    {
      "source": "memory-wall",
      "target": "tinyml",
      "type": "prerequisite"
    },
    {
      "source": "fleet-orchestration",
      "target": "topology-aware-placement",
      "type": "prerequisite"
    },
    {
      "source": "graph-compilation",
      "target": "torchdynamo",
      "type": "prerequisite"
    },
    {
      "source": "bisection-bandwidth",
      "target": "torus-topology",
      "type": "prerequisite"
    },
    {
      "source": "idempotent-transformations",
      "target": "transformation-lineage",
      "type": "prerequisite"
    },
    {
      "source": "fault-tolerance",
      "target": "transient-failures",
      "type": "prerequisite"
    },
    {
      "source": "roce",
      "target": "ultra-ethernet",
      "type": "prerequisite"
    },
    {
      "source": "custom-asic",
      "target": "wafer-scale-engine",
      "type": "prerequisite"
    },
    {
      "source": "checkpointing",
      "target": "warm-restart",
      "type": "prerequisite"
    },
    {
      "source": "simt-execution-model",
      "target": "warp-divergence",
      "type": "prerequisite"
    },
    {
      "source": "dataflow-optimization",
      "target": "weight-stationary",
      "type": "prerequisite"
    },
    {
      "source": "checkpointing",
      "target": "young-daly-formula",
      "type": "prerequisite"
    },
    {
      "source": "failure-rate",
      "target": "young-daly-formula",
      "type": "prerequisite"
    },
    {
      "source": "application-embedded-serving",
      "target": "zero-copy-inference",
      "type": "prerequisite"
    },
    {
      "source": "serialization-overhead",
      "target": "zero-copy-serialization",
      "type": "prerequisite"
    },
    {
      "source": "3d-parallelism",
      "target": "zero-optimizations",
      "type": "prerequisite"
    },
    {
      "source": "gradient-accumulation",
      "target": "data-parallelism",
      "type": "prerequisite"
    },
    {
      "source": "allreduce",
      "target": "data-parallelism",
      "type": "prerequisite"
    },
    {
      "source": "io-aware-algorithm-design",
      "target": "flash-attention",
      "type": "prerequisite"
    },
    {
      "source": "gpu-memory-hierarchy",
      "target": "hbm",
      "type": "prerequisite"
    },
    {
      "source": "operator-fusion",
      "target": "tensorrt",
      "type": "prerequisite"
    },
    {
      "source": "model-quantization",
      "target": "tensorrt",
      "type": "prerequisite"
    },
    {
      "source": "computational-graph",
      "target": "onnx-runtime",
      "type": "prerequisite"
    },
    {
      "source": "activation-memory-footprint",
      "target": "activation-logic-unit-cost",
      "type": "prerequisite"
    },
    {
      "source": "stationarity-assumption",
      "target": "adversarial-debiasing",
      "type": "prerequisite"
    },
    {
      "source": "model-memorization",
      "target": "adversarial-inputs",
      "type": "prerequisite"
    },
    {
      "source": "stationarity-assumption",
      "target": "alignment-gap",
      "type": "prerequisite"
    },
    {
      "source": "bisection-bandwidth",
      "target": "allgather",
      "type": "prerequisite"
    },
    {
      "source": "data-gravity",
      "target": "amdahl-s-law",
      "type": "prerequisite"
    },
    {
      "source": "batch-ingestion",
      "target": "amdahl-s-law-for-data",
      "type": "prerequisite"
    },
    {
      "source": "confidence-calibration",
      "target": "approximate-model-theft",
      "type": "prerequisite"
    },
    {
      "source": "matrix-operations",
      "target": "bandwidth-taper",
      "type": "prerequisite"
    },
    {
      "source": "sutva-violations",
      "target": "blue-green-deployment",
      "type": "prerequisite"
    },
    {
      "source": "computational-graph",
      "target": "caching-allocators",
      "type": "prerequisite"
    },
    {
      "source": "cmos-power-equation",
      "target": "cascade-inference",
      "type": "prerequisite"
    },
    {
      "source": "batch-ingestion",
      "target": "circuit-breaker-pattern",
      "type": "prerequisite"
    },
    {
      "source": "bisection-bandwidth",
      "target": "co-packaged-optics",
      "type": "prerequisite"
    },
    {
      "source": "bisection-bandwidth",
      "target": "compute-express-link",
      "type": "prerequisite"
    },
    {
      "source": "application-embedded-serving",
      "target": "constant-folding",
      "type": "prerequisite"
    },
    {
      "source": "data-cascade",
      "target": "continuous-deployment",
      "type": "prerequisite"
    },
    {
      "source": "confidence-calibration",
      "target": "counterfeit-hardware",
      "type": "prerequisite"
    },
    {
      "source": "application-embedded-serving",
      "target": "cpu-vectorization",
      "type": "prerequisite"
    },
    {
      "source": "activation-memory-footprint",
      "target": "credit-assignment-problem",
      "type": "prerequisite"
    },
    {
      "source": "model-memorization",
      "target": "data-anonymization-layer",
      "type": "prerequisite"
    },
    {
      "source": "data-cascade",
      "target": "data-lineage",
      "type": "prerequisite"
    },
    {
      "source": "stationarity-assumption",
      "target": "data-lineage-tracking",
      "type": "prerequisite"
    },
    {
      "source": "batch-ingestion",
      "target": "dataset-compilation",
      "type": "prerequisite"
    },
    {
      "source": "model-memorization",
      "target": "datasheets-for-datasets",
      "type": "prerequisite"
    },
    {
      "source": "sutva-violations",
      "target": "declarative-deployment-api",
      "type": "prerequisite"
    },
    {
      "source": "activation-memory-footprint",
      "target": "dense-connectivity-scaling",
      "type": "prerequisite"
    },
    {
      "source": "bisection-bandwidth",
      "target": "direct-attach-copper",
      "type": "prerequisite"
    },
    {
      "source": "activation-memory-footprint",
      "target": "double-descent-phenomenon",
      "type": "prerequisite"
    },
    {
      "source": "cmos-power-equation",
      "target": "duty-cycle-accounting",
      "type": "prerequisite"
    },
    {
      "source": "cuda-streams",
      "target": "efficiency-frontier",
      "type": "prerequisite"
    },
    {
      "source": "bisection-bandwidth",
      "target": "elephant-flows",
      "type": "prerequisite"
    },
    {
      "source": "cmos-power-equation",
      "target": "event-driven-activation",
      "type": "prerequisite"
    },
    {
      "source": "confidence-calibration",
      "target": "exact-model-theft",
      "type": "prerequisite"
    },
    {
      "source": "stationarity-assumption",
      "target": "fairness-accuracy-pareto-frontier",
      "type": "prerequisite"
    },
    {
      "source": "model-memorization",
      "target": "fairness-aware-training",
      "type": "prerequisite"
    },
    {
      "source": "kv-cache",
      "target": "feature-parallel-batching",
      "type": "prerequisite"
    },
    {
      "source": "adversarial-attacks",
      "target": "graceful-degradation",
      "type": "prerequisite"
    },
    {
      "source": "on-device-learning",
      "target": "gradient-checkpointing",
      "type": "prerequisite"
    },
    {
      "source": "cmos-power-equation",
      "target": "hardware-power-counters",
      "type": "prerequisite"
    },
    {
      "source": "matrix-operations",
      "target": "hardware-software-co-design",
      "type": "prerequisite"
    },
    {
      "source": "confidence-calibration",
      "target": "homomorphic-encryption",
      "type": "prerequisite"
    },
    {
      "source": "bisection-bandwidth",
      "target": "hpcc",
      "type": "prerequisite"
    },
    {
      "source": "stationarity-assumption",
      "target": "immutable-audit-trails",
      "type": "prerequisite"
    },
    {
      "source": "dense-matrix-multiplication",
      "target": "inductive-bias",
      "type": "prerequisite"
    },
    {
      "source": "adversarial-attacks",
      "target": "influence-functions",
      "type": "prerequisite"
    },
    {
      "source": "sutva-violations",
      "target": "interleaving-experiments",
      "type": "prerequisite"
    },
    {
      "source": "stationarity-assumption",
      "target": "invariance-testing",
      "type": "prerequisite"
    },
    {
      "source": "cmos-power-equation",
      "target": "knowledge-distillation",
      "type": "prerequisite"
    },
    {
      "source": "bisection-bandwidth",
      "target": "law-of-distributed-efficiency",
      "type": "prerequisite"
    },
    {
      "source": "confidence-calibration",
      "target": "leaky-interfaces",
      "type": "prerequisite"
    },
    {
      "source": "dense-matrix-multiplication",
      "target": "learnability-gap",
      "type": "prerequisite"
    },
    {
      "source": "kv-cache",
      "target": "little-s-law",
      "type": "prerequisite"
    },
    {
      "source": "application-embedded-serving",
      "target": "load-balancer",
      "type": "prerequisite"
    },
    {
      "source": "kv-cache",
      "target": "m-g-c-k-queue-model",
      "type": "prerequisite"
    },
    {
      "source": "model-memorization",
      "target": "machine-unlearning",
      "type": "prerequisite"
    },
    {
      "source": "application-embedded-serving",
      "target": "memory-planning",
      "type": "prerequisite"
    },
    {
      "source": "hardware-memory-hierarchy",
      "target": "metric-cardinality",
      "type": "prerequisite"
    },
    {
      "source": "bisection-bandwidth",
      "target": "mice-flows",
      "type": "prerequisite"
    },
    {
      "source": "matrix-operations",
      "target": "mixed-precision-computing",
      "type": "prerequisite"
    },
    {
      "source": "cuda-streams",
      "target": "mixture-of-experts",
      "type": "prerequisite"
    },
    {
      "source": "confidence-calibration",
      "target": "model-inversion-attack",
      "type": "prerequisite"
    },
    {
      "source": "cmos-power-equation",
      "target": "model-pruning",
      "type": "prerequisite"
    },
    {
      "source": "data-cascade",
      "target": "multi-scale-feedback",
      "type": "prerequisite"
    },
    {
      "source": "sutva-violations",
      "target": "multiple-testing-correction",
      "type": "prerequisite"
    },
    {
      "source": "application-embedded-serving",
      "target": "networked-serving",
      "type": "prerequisite"
    },
    {
      "source": "data-gravity",
      "target": "neural-processing-unit",
      "type": "prerequisite"
    },
    {
      "source": "confidence-calibration",
      "target": "offensive-machine-learning",
      "type": "prerequisite"
    },
    {
      "source": "cuda-streams",
      "target": "paged-kv-cache-management",
      "type": "prerequisite"
    },
    {
      "source": "data-gravity",
      "target": "pipelined-execution",
      "type": "prerequisite"
    },
    {
      "source": "cmos-power-equation",
      "target": "power-density-ceiling",
      "type": "prerequisite"
    },
    {
      "source": "cmos-power-equation",
      "target": "power-usage-effectiveness",
      "type": "prerequisite"
    },
    {
      "source": "kv-cache",
      "target": "preemptive-scheduling",
      "type": "prerequisite"
    },
    {
      "source": "bisection-bandwidth",
      "target": "quality-of-service",
      "type": "prerequisite"
    },
    {
      "source": "adversarial-attacks",
      "target": "robust-optimization",
      "type": "prerequisite"
    },
    {
      "source": "stationarity-assumption",
      "target": "role-based-access-control",
      "type": "prerequisite"
    },
    {
      "source": "confidence-calibration",
      "target": "secure-multiparty-computation",
      "type": "prerequisite"
    },
    {
      "source": "dense-matrix-multiplication",
      "target": "sequential-dependency",
      "type": "prerequisite"
    },
    {
      "source": "bisection-bandwidth",
      "target": "serializer-deserializer",
      "type": "prerequisite"
    },
    {
      "source": "sutva-violations",
      "target": "serving-autoscaling",
      "type": "prerequisite"
    },
    {
      "source": "kv-cache",
      "target": "serving-cost-dominance-law",
      "type": "prerequisite"
    },
    {
      "source": "kv-cache",
      "target": "serving-tax",
      "type": "prerequisite"
    },
    {
      "source": "hardware-memory-hierarchy",
      "target": "shadow-deployment",
      "type": "prerequisite"
    },
    {
      "source": "dense-matrix-multiplication",
      "target": "skip-connections",
      "type": "prerequisite"
    },
    {
      "source": "data-cascade",
      "target": "spatial-and-temporal-locality",
      "type": "prerequisite"
    },
    {
      "source": "matrix-operations",
      "target": "special-function-units",
      "type": "prerequisite"
    },
    {
      "source": "cuda-streams",
      "target": "speculative-decoding",
      "type": "prerequisite"
    },
    {
      "source": "confidence-calibration",
      "target": "speculative-execution-flaws",
      "type": "prerequisite"
    },
    {
      "source": "data-cascade",
      "target": "store-and-forward-architecture",
      "type": "prerequisite"
    },
    {
      "source": "batch-ingestion",
      "target": "stream-ingestion",
      "type": "prerequisite"
    },
    {
      "source": "matrix-operations",
      "target": "structured-sparsity",
      "type": "prerequisite"
    },
    {
      "source": "confidence-calibration",
      "target": "synthetic-data-generation",
      "type": "prerequisite"
    },
    {
      "source": "sutva-violations",
      "target": "telemetry-aggregation",
      "type": "prerequisite"
    },
    {
      "source": "kv-cache",
      "target": "test-time-compute-scaling",
      "type": "prerequisite"
    },
    {
      "source": "data-cascade",
      "target": "tiered-storage",
      "type": "prerequisite"
    },
    {
      "source": "application-embedded-serving",
      "target": "timeline-profiling",
      "type": "prerequisite"
    },
    {
      "source": "bisection-bandwidth",
      "target": "top-of-rack-switch",
      "type": "prerequisite"
    },
    {
      "source": "stationarity-assumption",
      "target": "total-cost-of-ownership",
      "type": "prerequisite"
    },
    {
      "source": "data-gravity",
      "target": "train-serve-split",
      "type": "prerequisite"
    },
    {
      "source": "cmos-power-equation",
      "target": "transient-loads",
      "type": "prerequisite"
    },
    {
      "source": "dense-matrix-multiplication",
      "target": "translation-equivariance",
      "type": "prerequisite"
    },
    {
      "source": "activation-memory-footprint",
      "target": "vanishing-gradient-problem",
      "type": "prerequisite"
    },
    {
      "source": "application-embedded-serving",
      "target": "zero-copy-loading",
      "type": "prerequisite"
    },
    {
      "source": "bisection-bandwidth",
      "target": "nvlink",
      "type": "prerequisite"
    },
    {
      "source": "accelerator-spectrum",
      "target": "pcie",
      "type": "prerequisite"
    },
    {
      "source": "activation-memory-footprint",
      "target": "memory-bandwidth",
      "type": "prerequisite"
    },
    {
      "source": "on-device-learning",
      "target": "coreml",
      "type": "prerequisite"
    },
    {
      "source": "on-device-learning",
      "target": "tflite",
      "type": "prerequisite"
    },
    {
      "source": "dense-matrix-multiplication",
      "target": "batch-normalization",
      "type": "prerequisite"
    },
    {
      "source": "dense-matrix-multiplication",
      "target": "layer-normalization",
      "type": "prerequisite"
    },
    {
      "source": "dense-matrix-multiplication",
      "target": "transformer-architecture",
      "type": "prerequisite"
    },
    {
      "source": "dense-matrix-multiplication",
      "target": "convolutional-neural-network",
      "type": "prerequisite"
    },
    {
      "source": "matrix-operations",
      "target": "sram",
      "type": "prerequisite"
    },
    {
      "source": "matrix-operations",
      "target": "cache-hierarchy",
      "type": "prerequisite"
    },
    {
      "source": "fleet-orchestration",
      "target": "load-balancing",
      "type": "prerequisite"
    },
    {
      "source": "hardware-memory-hierarchy",
      "target": "a-b-testing",
      "type": "prerequisite"
    },
    {
      "source": "data-benchmarking",
      "target": "mlperf",
      "type": "prerequisite"
    },
    {
      "source": "gradient-accumulation",
      "target": "low-rank-adaptation",
      "type": "prerequisite"
    },
    {
      "source": "mixed-precision-training",
      "target": "low-rank-adaptation",
      "type": "prerequisite"
    },
    {
      "source": "data-parallelism",
      "target": "fully-sharded-data-parallel",
      "type": "prerequisite"
    },
    {
      "source": "model-sharding",
      "target": "fully-sharded-data-parallel",
      "type": "prerequisite"
    },
    {
      "source": "allreduce",
      "target": "fully-sharded-data-parallel",
      "type": "prerequisite"
    },
    {
      "source": "flash-attention",
      "target": "flash-decoding",
      "type": "prerequisite"
    },
    {
      "source": "autoregressive-kv-cache",
      "target": "flash-decoding",
      "type": "prerequisite"
    },
    {
      "source": "memory-bound-workload",
      "target": "flash-decoding",
      "type": "prerequisite"
    },
    {
      "source": "continuous-batching",
      "target": "chunked-prefill",
      "type": "prerequisite"
    },
    {
      "source": "preemptive-scheduling",
      "target": "chunked-prefill",
      "type": "prerequisite"
    },
    {
      "source": "grouped-query-attention",
      "target": "multi-query-attention",
      "type": "prerequisite"
    },
    {
      "source": "autoregressive-kv-cache",
      "target": "multi-query-attention",
      "type": "prerequisite"
    },
    {
      "source": "quadratic-attention-bottleneck",
      "target": "sliding-window-attention",
      "type": "prerequisite"
    },
    {
      "source": "flash-attention",
      "target": "sliding-window-attention",
      "type": "prerequisite"
    },
    {
      "source": "quadratic-attention-bottleneck",
      "target": "sparse-attention",
      "type": "prerequisite"
    },
    {
      "source": "activation-sparsity",
      "target": "sparse-attention",
      "type": "prerequisite"
    },
    {
      "source": "vanishing-gradient-problem",
      "target": "gradient-clipping",
      "type": "prerequisite"
    },
    {
      "source": "mixed-precision-training",
      "target": "gradient-clipping",
      "type": "prerequisite"
    },
    {
      "source": "mixed-precision-training",
      "target": "dynamic-loss-scaling",
      "type": "prerequisite"
    },
    {
      "source": "gradient-clipping",
      "target": "dynamic-loss-scaling",
      "type": "prerequisite"
    },
    {
      "source": "gradient-accumulation",
      "target": "learning-rate-scheduling",
      "type": "prerequisite"
    },
    {
      "source": "low-rank-adaptation",
      "target": "adapter-layers",
      "type": "prerequisite"
    },
    {
      "source": "knowledge-distillation",
      "target": "adapter-layers",
      "type": "prerequisite"
    },
    {
      "source": "bin-packing-heuristics",
      "target": "gpu-virtualization",
      "type": "prerequisite"
    },
    {
      "source": "capacity-planning",
      "target": "gpu-virtualization",
      "type": "prerequisite"
    },
    {
      "source": "memory-wall",
      "target": "cxl-memory-expansion",
      "type": "prerequisite"
    },
    {
      "source": "cache-hierarchy",
      "target": "cxl-memory-expansion",
      "type": "prerequisite"
    },
    {
      "source": "tiered-storage",
      "target": "cxl-memory-expansion",
      "type": "prerequisite"
    },
    {
      "source": "transformer-architecture",
      "target": "tokenization",
      "type": "prerequisite"
    },
    {
      "source": "adversarial-inputs",
      "target": "red-teaming",
      "type": "prerequisite"
    },
    {
      "source": "certified-robustness",
      "target": "red-teaming",
      "type": "prerequisite"
    },
    {
      "source": "red-teaming",
      "target": "inference-guardrails",
      "type": "prerequisite"
    },
    {
      "source": "service-level-objective",
      "target": "inference-guardrails",
      "type": "prerequisite"
    },
    {
      "source": "load-balancer",
      "target": "request-routing",
      "type": "prerequisite"
    },
    {
      "source": "preemptive-scheduling",
      "target": "request-routing",
      "type": "prerequisite"
    },
    {
      "source": "capacity-planning",
      "target": "request-routing",
      "type": "prerequisite"
    },
    {
      "source": "autoregressive-kv-cache",
      "target": "token-budget",
      "type": "prerequisite"
    },
    {
      "source": "cost-per-inference",
      "target": "token-budget",
      "type": "prerequisite"
    },
    {
      "source": "data-parallelism",
      "target": "deepspeed-zero",
      "type": "prerequisite"
    },
    {
      "source": "fully-sharded-data-parallel",
      "target": "deepspeed-zero",
      "type": "prerequisite"
    },
    {
      "source": "gradient-accumulation",
      "target": "deepspeed-zero",
      "type": "prerequisite"
    },
    {
      "source": "tensor-parallelism",
      "target": "megatron-parallelism",
      "type": "prerequisite"
    },
    {
      "source": "allreduce",
      "target": "megatron-parallelism",
      "type": "prerequisite"
    },
    {
      "source": "3d-parallelism",
      "target": "megatron-parallelism",
      "type": "prerequisite"
    },
    {
      "source": "block-wise-quantization",
      "target": "gptq-quantization",
      "type": "prerequisite"
    },
    {
      "source": "model-quantization",
      "target": "gptq-quantization",
      "type": "prerequisite"
    },
    {
      "source": "block-wise-quantization",
      "target": "awq-quantization",
      "type": "prerequisite"
    },
    {
      "source": "model-quantization",
      "target": "awq-quantization",
      "type": "prerequisite"
    },
    {
      "source": "activation-sparsity",
      "target": "awq-quantization",
      "type": "prerequisite"
    },
    {
      "source": "low-rank-adaptation",
      "target": "qlora",
      "type": "prerequisite"
    },
    {
      "source": "block-wise-quantization",
      "target": "qlora",
      "type": "prerequisite"
    },
    {
      "source": "model-parallelism",
      "target": "rlhf-infrastructure",
      "type": "prerequisite"
    },
    {
      "source": "fully-sharded-data-parallel",
      "target": "rlhf-infrastructure",
      "type": "prerequisite"
    },
    {
      "source": "capacity-planning",
      "target": "rlhf-infrastructure",
      "type": "prerequisite"
    },
    {
      "source": "gpu-virtualization",
      "target": "container-gpu-scheduling",
      "type": "prerequisite"
    },
    {
      "source": "bin-packing-heuristics",
      "target": "container-gpu-scheduling",
      "type": "prerequisite"
    },
    {
      "source": "topology-aware-placement",
      "target": "container-gpu-scheduling",
      "type": "prerequisite"
    },
    {
      "source": "accelerator-spectrum",
      "target": "inference-accelerator-selection",
      "type": "prerequisite"
    },
    {
      "source": "cost-per-inference",
      "target": "inference-accelerator-selection",
      "type": "prerequisite"
    },
    {
      "source": "total-cost-of-ownership",
      "target": "inference-accelerator-selection",
      "type": "prerequisite"
    },
    {
      "source": "system-benchmarking",
      "target": "benchmarking-granularity",
      "type": "prerequisite"
    },
    {
      "source": "system-benchmarking",
      "target": "benchmark-harness",
      "type": "prerequisite"
    },
    {
      "source": "mlperf",
      "target": "benchmark-run-rules",
      "type": "prerequisite"
    },
    {
      "source": "data-gravity",
      "target": "information-entropy",
      "type": "prerequisite"
    },
    {
      "source": "schema-validation",
      "target": "data-quality-as-code",
      "type": "prerequisite"
    },
    {
      "source": "dataset-compilation",
      "target": "label-consensus",
      "type": "prerequisite"
    },
    {
      "source": "dataset-compilation",
      "target": "ai-assisted-labeling",
      "type": "prerequisite"
    },
    {
      "source": "arithmetic-intensity",
      "target": "energy-roofline-model-arithmetic-intensity",
      "type": "prerequisite"
    },
    {
      "source": "energy-per-byte",
      "target": "energy-roofline-model-arithmetic-intensity",
      "type": "prerequisite"
    },
    {
      "source": "energy-per-operation",
      "target": "energy-roofline-model-arithmetic-intensity",
      "type": "prerequisite"
    },
    {
      "source": "energy-roofline-model-arithmetic-intensity",
      "target": "energy-roofline-model-ridge-point",
      "type": "prerequisite"
    },
    {
      "source": "energy-roofline-model-ridge-point",
      "target": "energy-roofline-model-hardware",
      "type": "prerequisite"
    },
    {
      "source": "energy-roofline-model-ridge-point",
      "target": "energy-roofline-model-workloads",
      "type": "prerequisite"
    },
    {
      "source": "on-device-learning",
      "target": "federated-learning-foundations",
      "type": "prerequisite"
    },
    {
      "source": "federated-learning-foundations",
      "target": "federated-learning-economics",
      "type": "prerequisite"
    },
    {
      "source": "federated-learning-foundations",
      "target": "federated-learning-power",
      "type": "prerequisite"
    },
    {
      "source": "federated-learning-foundations",
      "target": "federated-learning-communication",
      "type": "prerequisite"
    },
    {
      "source": "federated-learning-foundations",
      "target": "federated-learning-privacy",
      "type": "prerequisite"
    },
    {
      "source": "gpu-memory-hierarchy-architecture",
      "target": "gpu-memory-hierarchy-capacity",
      "type": "prerequisite"
    },
    {
      "source": "gpu-memory-hierarchy-architecture",
      "target": "gpu-memory-hierarchy-bandwidth-latency",
      "type": "prerequisite"
    },
    {
      "source": "gpu-memory-hierarchy-capacity",
      "target": "gpu-memory-hierarchy-optimization",
      "type": "prerequisite"
    },
    {
      "source": "gpu-memory-hierarchy-bandwidth-latency",
      "target": "gpu-memory-hierarchy-optimization",
      "type": "prerequisite"
    },
    {
      "source": "application-embedded-serving-memory-footprint",
      "target": "application-embedded-serving-edge-devices",
      "type": "prerequisite"
    },
    {
      "source": "application-embedded-serving-memory-footprint",
      "target": "application-embedded-serving-cloud-scale",
      "type": "prerequisite"
    },
    {
      "source": "application-embedded-serving-batching-latency",
      "target": "application-embedded-serving-cloud-scale",
      "type": "prerequisite"
    },
    {
      "source": "application-embedded-serving-cloud-scale",
      "target": "application-embedded-serving-ops-hardware",
      "type": "prerequisite"
    },
    {
      "source": "arithmetic-intensity",
      "target": "compute-bound-workload",
      "type": "prerequisite"
    },
    {
      "source": "arithmetic-intensity",
      "target": "memory-bound-workload",
      "type": "prerequisite"
    },
    {
      "source": "autoregressive-kv-cache",
      "target": "dynamic-attention-routing",
      "type": "prerequisite"
    },
    {
      "source": "covariate-shift",
      "target": "population-stability-index",
      "type": "prerequisite"
    },
    {
      "source": "distribution-shift",
      "target": "silent-model-failures",
      "type": "prerequisite"
    },
    {
      "source": "data-striping",
      "target": "parallel-file-system",
      "type": "prerequisite"
    },
    {
      "source": "processing-elements",
      "target": "accelerator-spectrum",
      "type": "prerequisite"
    },
    {
      "source": "computational-graph",
      "target": "activation-memory",
      "type": "prerequisite"
    },
    {
      "source": "automatic-differentiation",
      "target": "activation-memory",
      "type": "prerequisite"
    },
    {
      "source": "activation-memory",
      "target": "activation-memory-footprint",
      "type": "prerequisite"
    },
    {
      "source": "mini-batch-processing",
      "target": "activation-memory-footprint",
      "type": "prerequisite"
    },
    {
      "source": "adversarial-inputs",
      "target": "adversarial-attacks",
      "type": "prerequisite"
    },
    {
      "source": "offensive-machine-learning",
      "target": "adversarial-attacks",
      "type": "prerequisite"
    },
    {
      "source": "inference-server",
      "target": "application-embedded-serving",
      "type": "prerequisite"
    },
    {
      "source": "tinyml",
      "target": "bare-metal-serving",
      "type": "prerequisite"
    },
    {
      "source": "leaky-interfaces",
      "target": "boundary-erosion",
      "type": "prerequisite"
    },
    {
      "source": "stage-interface-contracts",
      "target": "boundary-erosion",
      "type": "prerequisite"
    },
    {
      "source": "gradient-synchronization",
      "target": "bulk-synchronous-parallel",
      "type": "prerequisite"
    },
    {
      "source": "straggler-mitigation",
      "target": "bulk-synchronous-parallel",
      "type": "prerequisite"
    },
    {
      "source": "progressive-deployment",
      "target": "canary-deployment",
      "type": "prerequisite"
    },
    {
      "source": "service-level-objective",
      "target": "canary-deployment",
      "type": "prerequisite"
    },
    {
      "source": "matrix-operations",
      "target": "computational-graph",
      "type": "prerequisite"
    },
    {
      "source": "tensor-abstraction",
      "target": "computational-graph",
      "type": "prerequisite"
    },
    {
      "source": "system-benchmarking",
      "target": "micro-benchmarking",
      "type": "prerequisite"
    },
    {
      "source": "mlperf",
      "target": "time-to-train",
      "type": "prerequisite"
    },
    {
      "source": "system-benchmarking",
      "target": "time-to-train",
      "type": "prerequisite"
    },
    {
      "source": "dataset-compilation",
      "target": "inter-annotator-agreement",
      "type": "prerequisite"
    },
    {
      "source": "static-computation-graphs",
      "target": "micro-runtimes",
      "type": "prerequisite"
    },
    {
      "source": "tensor-abstraction",
      "target": "micro-runtimes",
      "type": "prerequisite"
    },
    {
      "source": "tensor-abstraction",
      "target": "neural-module-abstraction",
      "type": "prerequisite"
    },
    {
      "source": "computational-graph",
      "target": "neural-module-abstraction",
      "type": "prerequisite"
    },
    {
      "source": "computational-graph",
      "target": "mode-dependent-behavior",
      "type": "prerequisite"
    },
    {
      "source": "computational-graph",
      "target": "functional-transformations",
      "type": "prerequisite"
    },
    {
      "source": "automatic-differentiation",
      "target": "functional-transformations",
      "type": "prerequisite"
    },
    {
      "source": "hardware-software-co-design",
      "target": "domain-specific-architecture",
      "type": "prerequisite"
    },
    {
      "source": "sram",
      "target": "memory-tiling",
      "type": "prerequisite"
    },
    {
      "source": "matrix-operations",
      "target": "memory-tiling",
      "type": "prerequisite"
    },
    {
      "source": "memory-wall",
      "target": "memory-tiling",
      "type": "prerequisite"
    },
    {
      "source": "memory-bandwidth",
      "target": "hardware-balance",
      "type": "prerequisite"
    },
    {
      "source": "roofline-model",
      "target": "hardware-balance",
      "type": "prerequisite"
    },
    {
      "source": "silent-model-failures",
      "target": "ml-test-score",
      "type": "prerequisite"
    },
    {
      "source": "boundary-erosion",
      "target": "ml-test-score",
      "type": "prerequisite"
    },
    {
      "source": "continuous-training-pipelines",
      "target": "mlops-maturity-model",
      "type": "prerequisite"
    },
    {
      "source": "model-registry",
      "target": "mlops-maturity-model",
      "type": "prerequisite"
    },
    {
      "source": "boundary-erosion",
      "target": "pipeline-jungles",
      "type": "prerequisite"
    },
    {
      "source": "boundary-erosion",
      "target": "glue-code-antipattern",
      "type": "prerequisite"
    },
    {
      "source": "training-serving-skew",
      "target": "undeclared-data-dependencies",
      "type": "prerequisite"
    },
    {
      "source": "silent-model-failures",
      "target": "undeclared-data-dependencies",
      "type": "prerequisite"
    },
    {
      "source": "distribution-shift",
      "target": "online-evaluation-strategies",
      "type": "prerequisite"
    },
    {
      "source": "data-lineage",
      "target": "system-artifact-provenance",
      "type": "prerequisite"
    },
    {
      "source": "constraint-propagation",
      "target": "deployment-pareto-optimization",
      "type": "prerequisite"
    },
    {
      "source": "throughput",
      "target": "littles-law",
      "type": "prerequisite"
    },
    {
      "source": "latency",
      "target": "littles-law",
      "type": "prerequisite"
    },
    {
      "source": "capacity-planning",
      "target": "littles-law",
      "type": "prerequisite"
    },
    {
      "source": "inference-pipeline",
      "target": "request-pipelining",
      "type": "prerequisite"
    },
    {
      "source": "throughput",
      "target": "request-pipelining",
      "type": "prerequisite"
    },
    {
      "source": "tail-latency",
      "target": "tail-tolerant-serving",
      "type": "prerequisite"
    },
    {
      "source": "load-balancer",
      "target": "tail-tolerant-serving",
      "type": "prerequisite"
    },
    {
      "source": "inference-server",
      "target": "cold-start",
      "type": "prerequisite"
    },
    {
      "source": "latency",
      "target": "cold-start",
      "type": "prerequisite"
    },
    {
      "source": "inference-server",
      "target": "multi-model-serving",
      "type": "prerequisite"
    },
    {
      "source": "inference-server",
      "target": "model-caching",
      "type": "prerequisite"
    },
    {
      "source": "multi-model-serving",
      "target": "model-caching",
      "type": "prerequisite"
    },
    {
      "source": "transformer-architecture",
      "target": "positional-encoding",
      "type": "prerequisite"
    },
    {
      "source": "convolutional-neural-network",
      "target": "receptive-field",
      "type": "prerequisite"
    },
    {
      "source": "dense-connectivity-scaling",
      "target": "hierarchical-feature-learning",
      "type": "prerequisite"
    },
    {
      "source": "credit-assignment-problem",
      "target": "end-to-end-differentiability",
      "type": "prerequisite"
    },
    {
      "source": "multiply-accumulate-operations",
      "target": "backpropagation-compute-cost",
      "type": "prerequisite"
    },
    {
      "source": "credit-assignment-problem",
      "target": "backpropagation-compute-cost",
      "type": "prerequisite"
    },
    {
      "source": "matrix-multiplication-workloads",
      "target": "inference-pipeline-latency",
      "type": "prerequisite"
    },
    {
      "source": "alignment-gap",
      "target": "runaway-feedback-loops",
      "type": "prerequisite"
    },
    {
      "source": "distribution-shift",
      "target": "runaway-feedback-loops",
      "type": "prerequisite"
    },
    {
      "source": "distribution-shift",
      "target": "model-fallback-strategies",
      "type": "prerequisite"
    },
    {
      "source": "computational-intensity",
      "target": "backpropagation-compute-ratio",
      "type": "prerequisite"
    },
    {
      "source": "iron-law-of-training",
      "target": "training-physical-ceiling",
      "type": "prerequisite"
    },
    {
      "source": "staged-system-pipeline",
      "target": "training-physical-ceiling",
      "type": "prerequisite"
    },
    {
      "source": "alpha-beta-performance-model",
      "target": "logp-model",
      "type": "prerequisite"
    },
    {
      "source": "allreduce",
      "target": "tree-allreduce",
      "type": "prerequisite"
    },
    {
      "source": "allreduce",
      "target": "recursive-halving-doubling",
      "type": "prerequisite"
    },
    {
      "source": "allreduce",
      "target": "in-network-reduction",
      "type": "prerequisite"
    },
    {
      "source": "hierarchical-allreduce",
      "target": "topology-aware-routing",
      "type": "prerequisite"
    },
    {
      "source": "gradient-compression",
      "target": "compression-aware-optimization",
      "type": "prerequisite"
    },
    {
      "source": "pcie",
      "target": "high-speed-gpu-interconnect",
      "type": "prerequisite"
    },
    {
      "source": "memory-wall",
      "target": "high-speed-gpu-interconnect",
      "type": "prerequisite"
    },
    {
      "source": "pcie",
      "target": "non-uniform-memory-access",
      "type": "prerequisite"
    },
    {
      "source": "pcie",
      "target": "direct-storage-access",
      "type": "prerequisite"
    },
    {
      "source": "memory-wall",
      "target": "direct-storage-access",
      "type": "prerequisite"
    },
    {
      "source": "thermal-design-power",
      "target": "rack-power-density",
      "type": "prerequisite"
    },
    {
      "source": "thermal-design-power",
      "target": "direct-to-chip-liquid-cooling",
      "type": "prerequisite"
    },
    {
      "source": "thermal-design-power",
      "target": "infrastructure-tco",
      "type": "prerequisite"
    },
    {
      "source": "data-parallelism",
      "target": "stale-synchronous-parallel",
      "type": "prerequisite"
    },
    {
      "source": "data-parallelism",
      "target": "learning-rate-scaling-rules",
      "type": "prerequisite"
    },
    {
      "source": "data-parallelism",
      "target": "critical-batch-size",
      "type": "prerequisite"
    },
    {
      "source": "data-parallelism",
      "target": "parameter-server-architecture",
      "type": "prerequisite"
    },
    {
      "source": "model-flops-utilization",
      "target": "expert-parallelism",
      "type": "prerequisite"
    },
    {
      "source": "compute-bound",
      "target": "fp8-training",
      "type": "prerequisite"
    },
    {
      "source": "communication-bound",
      "target": "fp8-training",
      "type": "prerequisite"
    },
    {
      "source": "failure-rate",
      "target": "bathtub-curve",
      "type": "prerequisite"
    },
    {
      "source": "failure-rate",
      "target": "mean-time-between-failures",
      "type": "prerequisite"
    },
    {
      "source": "fault-model",
      "target": "permanent-faults",
      "type": "prerequisite"
    },
    {
      "source": "fault-model",
      "target": "intermittent-faults",
      "type": "prerequisite"
    },
    {
      "source": "distributed-checkpointing",
      "target": "checkpoint-storm",
      "type": "prerequisite"
    },
    {
      "source": "checkpointing",
      "target": "checkpoint-storm",
      "type": "prerequisite"
    },
    {
      "source": "alpha-beta-performance-model",
      "target": "five-level-networking-model",
      "type": "prerequisite"
    },
    {
      "source": "serializer-deserializer",
      "target": "link-budget",
      "type": "prerequisite"
    },
    {
      "source": "pam4",
      "target": "link-budget",
      "type": "prerequisite"
    },
    {
      "source": "top-of-rack-switch",
      "target": "failure-domain",
      "type": "prerequisite"
    },
    {
      "source": "gpudirect-rdma",
      "target": "gradient-bus",
      "type": "prerequisite"
    },
    {
      "source": "allreduce",
      "target": "gradient-bus",
      "type": "prerequisite"
    },
    {
      "source": "model-registry",
      "target": "ml-technical-debt",
      "type": "prerequisite"
    },
    {
      "source": "serving-autoscaling",
      "target": "ml-capacity-planning",
      "type": "prerequisite"
    },
    {
      "source": "model-registry",
      "target": "ensemble-management",
      "type": "prerequisite"
    },
    {
      "source": "dependency-aware-deployment",
      "target": "ensemble-management",
      "type": "prerequisite"
    },
    {
      "source": "model-registry",
      "target": "continuous-training-pipeline",
      "type": "prerequisite"
    },
    {
      "source": "model-registry",
      "target": "automated-validation-gates",
      "type": "prerequisite"
    },
    {
      "source": "telemetry-aggregation",
      "target": "hierarchical-monitoring",
      "type": "prerequisite"
    },
    {
      "source": "telemetry-aggregation",
      "target": "alert-fatigue-mitigation",
      "type": "prerequisite"
    },
    {
      "source": "efficiency-frontier",
      "target": "neural-architecture-search",
      "type": "prerequisite"
    },
    {
      "source": "efficiency-frontier",
      "target": "hardware-aware-nas",
      "type": "prerequisite"
    },
    {
      "source": "roofline-model",
      "target": "hardware-aware-nas",
      "type": "prerequisite"
    },
    {
      "source": "precision-engineering",
      "target": "automated-model-compression",
      "type": "prerequisite"
    },
    {
      "source": "efficiency-frontier",
      "target": "automated-model-compression",
      "type": "prerequisite"
    },
    {
      "source": "distribution-shift",
      "target": "system-feedback-loops",
      "type": "prerequisite"
    },
    {
      "source": "system-explainability",
      "target": "algorithmic-contestability",
      "type": "prerequisite"
    },
    {
      "source": "adversarial-inputs",
      "target": "llm-guardrails",
      "type": "prerequisite"
    },
    {
      "source": "membership-inference-attack",
      "target": "differential-privacy-mechanisms",
      "type": "prerequisite"
    },
    {
      "source": "distribution-shift",
      "target": "out-of-distribution-detection",
      "type": "prerequisite"
    },
    {
      "source": "adversarial-attacks",
      "target": "prompt-injection",
      "type": "prerequisite"
    },
    {
      "source": "layered-defense",
      "target": "threat-modeling-ml",
      "type": "prerequisite"
    },
    {
      "source": "runtime-input-validation",
      "target": "adversarial-evasion-attack",
      "type": "prerequisite"
    },
    {
      "source": "runtime-input-validation",
      "target": "llm-prompt-injection",
      "type": "prerequisite"
    },
    {
      "source": "layered-defense",
      "target": "ml-supply-chain-security",
      "type": "prerequisite"
    },
    {
      "source": "power-usage-effectiveness",
      "target": "water-usage-effectiveness",
      "type": "prerequisite"
    },
    {
      "source": "operational-carbon",
      "target": "grid-carbon-intensity",
      "type": "prerequisite"
    },
    {
      "source": "embodied-carbon",
      "target": "embodied-carbon-amortization",
      "type": "prerequisite"
    },
    {
      "source": "lifecycle-carbon-accounting",
      "target": "embodied-carbon-amortization",
      "type": "prerequisite"
    },
    {
      "source": "power-density-ceiling",
      "target": "divergent-energy-scaling",
      "type": "prerequisite"
    },
    {
      "source": "cmos-power-equation",
      "target": "divergent-energy-scaling",
      "type": "prerequisite"
    }
  ],
  "stats": {
    "total_concepts": 839,
    "total_edges": 746,
    "orphan_concepts": 0,
    "chapters_processed": 27,
    "version_date": "2026-03-24",
    "version_notes": "v4.2: Redistributed 729 overloaded concepts, assigned 438 empty taxonomy_concepts, deep math validation on all 5228 Qs"
  }
}