PowerDL – Unified In-Memory GPU Energy Profiling and Visualization Toolkit

PowerDL is a lightweight, framework-agnostic toolkit for GPU energy, power, and utilization profiling of deep learning workloads.
It supports PyTorch and TensorFlow, operates fully in-memory by default, and provides rich, publication-ready visual analytics with optional artifact export.

The toolkit is designed for:

• Interactive experimentation (no mandatory file I/O)
• Energy-aware deep learning research
• Reproducible benchmarking
• Software-oriented scientific contributions

---

Key Features

• Unified high-level API: profile_torch, profile_tf
• Fully in-memory profiling (no files required by default)
• Optional export for reproducibility
• High-resolution time-series sampling (NVML-based)
• Phase-aware analysis (training / inference / warmup / session)
• Batch-level and epoch-level energy breakdown
• Optional visualization layer
• Clean separation between measurement and plotting
• Research-ready structured outputs (CSV + JSON)

---

Installation

Minimal Core Install (No Visualization)

pip install powerdl

With Visualization Support

pip install powerdl[viz]

With PyTorch Examples

pip install powerdl[examples-torch]

With TensorFlow Examples

pip install powerdl[examples-tf]

---

Installation from Source (for running examples)

git clone https://github.com/your-repo/powerdl.git cd powerdl pip install -e .[viz]

---

Requirements

The following dependencies are required for full functionality:

• Python ≥ 3.9
• NVIDIA GPU
• NVIDIA driver with NVML support
• pandas
• matplotlib
• SciencePlots
• nvidia-ml-py

Optional (depending on framework):

• torch
• tensorflow

---

Running Example Scripts (IMPORTANT)

If you are using the repository version, install the package first:

pip install -e .

Then run examples from the project root:

python -m examples.run_torch_example

or

python -m examples.run_tf_example

Do NOT run scripts directly using:

python examples/run_torch_example.py

This may result in import errors because Python cannot locate the powerdl module.

Using the "-m" option ensures correct module resolution and avoids the need for manual sys.path modifications.

---

Quick Start – PyTorch

from powerdl.highlevel import profile_torch

with profile_torch( out_dir=None, interval_s=0.02, verbose=1 ) as prof:

prof.train(model, trainloader, optimizer, loss_fn, epochs=3)
prof.infer_tensor(model, batch_size=256, n_samples=20000)

rep = prof.report() rep.export("results/torch_run")

---

Quick Start – TensorFlow

from powerdl.highlevel import profile_tf

with profile_tf( out_dir=None, interval_s=0.02, verbose=1 ) as prof:

prof.fit(model, dataset, epochs=3)
prof.infer_keras(model, x_inf, batch_size=512)

rep = prof.report() rep.export("results/tf_run")

---

In-Memory First Design

PowerDL collects all measurements in memory first.

Collected signals:

• GPU power (W)
• GPU utilization (%)
• Memory utilization (%)
• Precise timestamps
• Phase markers (training, inference, epochs, warmup)

No files are written unless explicitly requested.

Optional export: rep.export("runs/experiment_export")

---

Report Object

rep = prof.report()

The unified Report object contains:

• samples_df – full time-series measurements
• marks_df – event and phase markers
• summary – structured scalar metrics

Visualization is lazy-loaded and requires matplotlib only when plotting is used.

---

Visualization System (Optional)

pip install powerdl[viz]

rep.list_figures()

---

Supported Figures

Time-Series:

• power_time
• gpu_util_time
• mem_util_time
• util_time_dual
• energy_rate
• power_derivative
• cumulative_energy

Distribution-Based:

• power_hist
• util_hist
• power_ecdf
• util_ecdf

Phase-Aware Statistics:

• power_boxplot_phase
• util_boxplot_phase
• phase_energy_bar
• phase_time_bar

Correlation & Density:

• power_util_scatter
• power_util_hexbin

---

Plotting API

Plot all figures:

rep.plot( all=True, out_dir="results/figs", show=False, smooth=5, shade_phases=True )

Plot selected figures:

rep.plot( figs=[ "power_time", "cumulative_energy", "power_util_hexbin", "phase_energy_bar" ], out_dir="results/figs", show=True )

Plot single figure:

rep.plot_one("power_ecdf", show=True)

---

Exported Artifacts

rep.export("runs/experiment_01")

Produces:

• summary.json
• samples.csv
• marks.csv
• batches.csv
• epochs.csv

---

Design Philosophy

PowerDL separates:

• Measurement (NVML sampling)
• Aggregation (energy integration and phase pairing)
• Visualization (registry-based plotting)

Focus:

• temporal analysis
• phase-aware energy behavior
• reproducible research

---

Typical Use Cases

• Green AI benchmarking
• Training vs inference energy comparison
• GPU saturation analysis
• Phase-wise energy profiling
• Deep learning energy efficiency studies

---

License

MIT License

---

Author

Yasin SANCAR
Computer Engineer
Energy-Aware Deep Learning Research