hongpingzhang/ecocompute-benchmark

EcoCompute — LLM Quantization Energy Benchmark

Real GPU power measurements (NVML, 10 Hz) for FP16 / INT8 / NF4 inference across 4 NVIDIA architectures (T4 / A800 / RTX 4090D / RTX 5090) and 360+ configurations in the 1.1B–14B parameter range, with WikiText-2 perplexity for accuracy trade-offs.

Files

• EcoCompute_Practitioner_Report_v1.0.pdf — 6-page practitioner report: key findings, decision tables, and operating-point recommendations.

(Raw CSVs and per-row metadata live in the GitHub mirror — see Links below.)

Headline findings

1. LLM.int8() default is an energy regression on RTX 4090D: +80%–+162% energy vs FP16, throughput −68% to −75%, perplexity essentially preserved.
2. NF4 has a GPU-specific crossover at ~3.2B (Turing) → ~5.2B (Blackwell). Below the crossover, NF4 adds 25%–55% energy. Above it, NF4 saves 15%–23%.

Citation

@misc{zhang2026llmenergy,
  author = {Zhang, Hongping},
  title  = {LLM Energy Benchmark: Real GPU Power Measurements for Quantized Inference},
  year   = {2026},
  doi    = {10.5281/zenodo.18900289},
  url    = {https://doi.org/10.5281/zenodo.18900289}
}

Links

• Discussion (HF Forums): https://discuss.huggingface.co/t/i-measured-360-configs-quantization-often-costs-energy-below-the-crossover-point/175979
• Code + raw CSVs: https://github.com/hongping-zh/ecocompute-ai
• Zenodo archive (DOI): https://doi.org/10.5281/zenodo.18900289
• Interactive comparison tool: https://hongping-zh.github.io/compare.html

License

Released under CC BY 4.0 — please cite when you use the data or report.