What does "GPU utilization" mean? Code and data

Code and data for the preprint:

Vilanakis, N. D. (2026). What does "GPU utilization" mean? Testing the 30% accuracy margin for training-compute estimates under the EU AI Act. Preprint. Zenodo. 10.5281/zenodo.22881088

Under Article 51(2) of the EU AI Act, a general-purpose AI model trained with more than 10²⁵ FLOP is presumed to carry systemic risk. The Commission's Guidelines (18 July 2025, Annex A.2) let providers estimate training compute with a hardware-based method, C = N × L × H × U, or an architecture-based method, C ≈ 6 × P × D, and accept any estimate that is accurate within 30%. This repository measures how much the undefined terms, "GPU utilisation" U and peak performance H, change the result.

Contents

Folder	What it is
flopcheck/	Reference implementation of both Annex A.2 methods. It reproduces the Commission's worked examples and has tests.
probe/	The H100 measurement script (h100_probe.py) and the raw results from one NVIDIA H100 PCIe, 16 September 2026.
analysis/	Scripts that turn the measurements into the paper's figures and tables, with their outputs in analysis/out/.
data/dos-group-gpu-power-benchmark/	Third-party benchmark data (Enskat, Wiesner & Kao, MIT licence), re-analysed in the paper. See its own README.md and SOURCE_COMMIT.txt.

Reproducing the results

Python 3.10 or newer.

flopcheck (no dependencies apart from pytest):

cd flopcheck
pip install -e ".[dev]"
pytest
python -m flopcheck.report

Figures 1–5 (needs pandas and matplotlib; run from the repository root):

pip install pandas matplotlib
python analysis/reanalysis_dos_group.py
python analysis/h100_probe_analysis.py

Script	Input	Output in analysis/out/
reanalysis_dos_group.py	data/dos-group-gpu-power-benchmark/	fig1_util_vs_mfu, fig2_gap_at_batch128, fig3_flop_count_methods, fp32_peak_report.csv, summary.json
h100_probe_analysis.py	probe/results/20260916T190612Z/	fig4_h100_counter_vs_mfu, fig5_h100_recomputation, h100_probe_summary.csv, h100_probe_summary.json

analysis/REANALYSIS.md and analysis/H100_PROBE.md describe the methods and the results in detail.

Re-running the H100 measurement needs an NVIDIA H100 and PyTorch 2.3 or newer. probe/README.md walks through it step by step on a rented cloud instance (about one hour).

cd probe
pip install -r requirements.txt
python h100_probe.py preflight
python h100_probe.py run

The H100 results

Folder	What it holds
probe/results/20260916T190612Z/	The run used in the paper: six configurations (sequence length 2,048 or 8,192; with and without activation checkpointing)
probe/results/preflight/	The 15-second preflight check

Each run folder contains env.json (GPU, driver and software versions), results.jsonl (one line per configuration), nvml_*.csv (raw NVML samples) and dcgm_*.log (raw DCGM output).

Citation

If you use this code or data, please cite the preprint above. GitHub's "Cite this repository" button uses CITATION.cff.

Licence

MIT, see LICENSE. The data in data/dos-group-gpu-power-benchmark/ keeps its own MIT licence, © Niklas Enskat, Philipp Wiesner and Odej Kao.