The Joule Point: an Energy-Optimal Operating Point for AI Inference

Alexander Apartsin (HIT) and Yehudit Aperstein (Afeka). Paper source, builds, dataset, measurement harnesses, figures, fleet simulation, and review artifacts.

Paper (HTML): docs/GPTEnergy.html (GitHub Pages serves it as the site index, with a PDF link at the top right). Paper (PDF/Word): docs/JoulePoint_1col.pdf (primary, single-column), docs/JoulePoint_2col.pdf, plus the matching .docx files. Dataset: ELF (Energy-Latency Frontier), archived at Zenodo, doi:10.5281/zenodo.22058568, CC-BY-4.0 (data) and MIT (code).

Headline

GPU board power follows P(θ) = P₀ + aθ^β over the operating point (GPU, power cap), so energy per inference is U-shaped with a minimum, the Joule Point, at 43 to 46 per cent of TDP on large GPUs. Capping to it cuts energy per inference by 29 to 31 per cent at an exactly priced cost (about 1.2x latency and the same factor in cards). Under load the Joule Point is nearly a per-card constant, so one static cap per card replaces the online per-job search prior systems run; a trace-driven fleet simulation over the measured curves serves equal work for 18 to 45 per cent less energy under a power budget.

Layout

docs/            GPTEnergy.html            the paper (built HTML, single source of truth)
                 GPTEnergy_{1col,2col}.{pdf,docx}   built deliverables
                 build_docx.py             HTML -> Word/PDF build
power_shaping/   build_gptenergy.py        writes docs/GPTEnergy.html
                 aws_*.py                  AWS measurement harnesses (power-cap and clock sweeps)
                 make_*.py                 figure builders
                 figures/                  built figures used by the paper
                 data/raw, data/processed  measured sweeps; elf_master.csv is the merged dataset
                 elf_release/              the Zenodo deposit package (README, dictionary, LICENSE)
                 energy_scheduler_sim.py   fleet simulation of Section 7
                 results/                  per-experiment JSON results
                 reviews/                  external review passes on the draft
reports/         dataset verification, novelty and related-work scouting, audit notes
experiments/     earlier pilot line (configuration-dependent energy rankings), with BUGFIXES.md
paper/           earlier draft lineage (greenmatch)
references/      cited external material

Reproducing

Every empirical figure and measurement-derived number in the paper is computed from the ELF data by the build scripts:

python power_shaping/build_gptenergy.py   # rebuilds docs/GPTEnergy.html
python docs/build_docx.py                 # rebuilds the Word/PDF deliverables

The power_shaping/aws_*.py harnesses reproduce the measurements themselves on rented AWS GPU instances (g6, g5, p4d, g4dn); see power_shaping/DATA_PROVENANCE.md.