Which EcoCompute artifact should I cite?

There are three artifacts and they do three different jobs. This page is the single authoritative list — what each one is, how they relate, and which DOI to use.

At a glance

Zenodo mints two DOIs per record: a concept DOI that always redirects to the latest version, and a version DOI that is frozen. Both appear across this site and the READMEs — they point at the same deposit, they are not competing artifacts.

How they relate

preprint (SSRN #6854700, Zenodo 21066652)
   |  states the claim and the NVML measurement protocol
   |
   +-- main dataset v1.1.0 (19647289 -> 19647290)  ......  SOURCE OF TRUTH for this site
   |      360+ configurations, 0.5B-14B, 4 architectures, FP16/NF4/INT8/FP8, n=2, CV < 2%
   |      supersedes  dataset v1 (18900288 -> 18900289, 113 configurations)
   |
   +-- RTX 4090 (Ada) deep dive (22019741 -> 22037483)  ..  SUPPLEMENTS v1.1.0, does not replace it
          1 card, produced by the open container (ecocompute-mlcube)
          session 2026-07-24: 15/15 measured, 0.5B-7B x FP16/NF4/INT8, energy only, n=1
          session 2026-08-19: 10 measured, 0.5B-7B x NF4/INT8, energy + paired perplexity, n=1
                              INT8 penalty 2.0-2.5x the July session's -> do not pool
          session 2026-08-20: the 5 INT8 configurations re-run 3x each, n=3
                              CV of the energy delta 0.6-3.9%, i.e. 30-50x smaller than that gap
          supersedes  21528102 -> 21528103 (its vs_fp16_pct column contradicted its energy column)

Concretely: supplements means the RTX 4090 run adds coverage v1.1.0 does not have (INT8 on Ada, and 7B on Ada) at a lower repeat count, so it is a deep dive rather than a replacement — and if you quote a figure from the RTX 4090 column, cite it; otherwise cite v1.1.0. Supersedes means v1 should no longer be cited: any mirror or skill listing that still says “113+ configurations” is describing v1.

Each artifact in full

@article{zhang_quantization_energy,
  title   = {Weight-Only Quantization Does Not Always Save Energy: An Empirical
             Study of LLM Inference Across NVIDIA GPU Platforms},
  author  = {Zhang, Hongping},
  year    = {2026},
  note    = {Preprint, SSRN abstract 6854700},
  doi     = {10.5281/zenodo.21066654},
  url     = {https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6854700}
}

@dataset{zhang_ecocompute_dataset_v1_1_0,
  title     = {EcoCompute: Energy Efficiency Benchmark for Quantized
               Language Models},
  author    = {Zhang, Hongping},
  year      = {2026},
  publisher = {Zenodo},
  version   = {1.1.0},
  doi       = {10.5281/zenodo.19647290},
  url       = {https://doi.org/10.5281/zenodo.19647290}
}

@dataset{zhang_ecocompute_rtx4090_2026,
  title     = {EcoCompute --- LLM Inference Energy on NVIDIA RTX 4090 (Ada)},
  author    = {Zhang, Hongping},
  year      = {2026},
  publisher = {Zenodo},
  note      = {Supplementary single-platform deep dive to dataset v1.1.0
               (10.5281/zenodo.19647290); direct NVML, FP16/NF4/INT8, 0.5B--7B;
               three sessions: energy only, energy with paired perplexity,
               and INT8 repeated 3x per size (CV 0.6--3.9%);
               supersedes 10.5281/zenodo.21528103},
  doi       = {10.5281/zenodo.22037483},
  url       = {https://doi.org/10.5281/zenodo.22037483}
}

Software, not data

• ecocompute-mlcube — the measurement container that produced the RTX 4090 run (what it does and how to run it)
• ecocompute-ai — dataset processing and analysis code
• quant-energy — this site, the estimator (estimate.js) and the optional REST API

Cite software by repository URL and commit; the DOIs above cover the paper and the data only.

@software{zhang_ecocompute_mlcube,
  title   = {EcoCompute energy MLCube: reproducible LLM inference energy measurement},
  author  = {Zhang, Hongping},
  year    = {2026},
  note    = {MLCube-compatible measurement container; cite the commit you ran},
  url     = {https://github.com/hongping-zh/ecocompute-mlcube}
}

Press & media kit

One sentence: EcoCompute is an open reproducibility project measuring how LLM quantization affects inference energy per token across GPUs and model sizes.

One paragraph: Weight-only quantization (NF4, INT8) is widely assumed to save energy because it saves memory. Direct NVML power measurements across four NVIDIA GPU architectures show that below a model-size crossover point it often costs energy instead. EcoCompute publishes those measurements, the fitted crossover curves, an interactive estimator, and — crucially — the MLCube-compatible container that produced them, so anyone with an NVIDIA GPU can reproduce or contradict the result. It is independent research, not a certified benchmark.

Please describe EcoCompute as independent research. It is not a certified benchmark result, and it carries no endorsement from any standards body.