ViT

ViT MFU Benchmark

Measures Model FLOP Utilization (MFU) for Vision Transformer training across three storage backends: LanceDB (OSS + Enterprise), raw S3 via Boto3, and Parquet on S3.

The benchmark uses synthetic JPEG images at a fixed resolution to isolate data pipeline throughput as the variable. All three scripts run the same ViT model with the same training loop, so the MFU difference is entirely a function of how fast each backend can feed batches.

Scripts

Script	Purpose
mfu_bench_fp16/bench.py	Training loop + all three storage backends
mfu_bench_fp16/dataloaders.py	LanceDB, Boto3, and Parquet DataLoader implementations
create_data.py	Generates the dataset and stages it in all three backends

Setup

Requires a CUDA GPU. Configure your AWS credentials and LanceDB connection details in each script (or via environment variables).

# Generate and stage data
uv run python examples/ViT/create_data.py

# Run benchmark (all three backends in one script)
uv run python examples/ViT/mfu_bench_fp16/bench.py

Configuration

Each benchmark script shares the same knobs at the top of the file:

• MODEL_NAME: "vit_h_14", "vit_l_16", or "vit_b_16"
• EXPECT_IMAGE_SIZE: input resolution, default (224, 224)
• BATCH_SIZE: images per training step
• WARMUP_STEPS / BENCH_STEPS: steps to discard / steps to time
• NUM_WORKERS: DataLoader workers
• PEAK_FLOPS: your GPU's peak bfloat16 FLOPS (default set for H100/H200)

Output

Each script prints throughput, achieved TFLOPS, and GPU MFU after the timed window.

H200 — vit_h_14 @ 224×224

--- Synthetic Pure-GPU Baseline (vit_h_14) / (224, 224) ---
batch_size=350, warmup=5, steps=50
Time Taken:     43.349 sec
Throughput:     403.70 images/sec
Achieved FLOPS: 405.221 TFLOPS
GPU MFU:        40.97%

--- LanceDB OSS Training (vit_h_14) / (224, 224) ---
batch_size=350, workers=8, steps=50
Time Taken:     47.597 sec
Throughput:     367.67 images/sec
Achieved FLOPS: 369.060 TFLOPS
GPU MFU:        37.32%

--- LanceDB Enterprise Training (vit_h_14) / (224, 224) ---
batch_size=350, workers=8, steps=50
Time Taken:     45.773 sec
Throughput:     382.32 images/sec
Achieved FLOPS: 383.762 TFLOPS
GPU MFU:        38.80%

--- Boto3 S3 Training (vit_h_14) / (224, 224) ---
batch_size=350, workers=8, steps=50
Time Taken:     137.291 sec
Throughput:     127.47 images/sec
Achieved FLOPS: 127.948 TFLOPS
GPU MFU:        12.94%

--- S3 Parquet Training (vit_h_14) / (224, 224) ---
batch_size=350, workers=8, steps=50
Time Taken:     72.393 sec
Throughput:     207.20 images/sec
Achieved FLOPS: 207.984 TFLOPS
GPU MFU:        21.03%