Which EcoCompute artifact should I cite?

1 · Preprint — the argument and the method

Title Weight-Only Quantization Does Not Always Save Energy: An Empirical Study of LLM Inference Across NVIDIA GPU Platforms

Where SSRN #6854700 · archived copy on Zenodo

Cite it for the crossover claim, the protocol, the framing

@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}
}

2 · Protocol v1.1 — the normative specification

Published 2026-09-25 · CC BY 4.0

Contents the measurement specification itself, extracted from the site's method layer into a standalone W3C-Note/IETF-RFC-style document: workload (single-stream decode, batch 1, 256 generated tokens), baseline (same-session same-card FP16), measurement source (NVML GPU-package power at ≥10 Hz), measurement window, report fields, compliance levels A/B/C, and the protocol ↔ schema ↔ dataset version table. A conforming submission MUST pass ecocompute-energy/1.2+ validation.

Cite it for the measurement method in any paper that follows it — cite this instead of the website's /method/ page, which is the reader's guide

@techreport{zhang_ecocompute_protocol_v1_1,
  title       = {{EcoCompute Protocol v1.1: A Micro-Benchmark
                  Specification for Quantization Energy in LLM Inference}},
  author      = {Zhang, Hongping},
  year        = {2026},
  month       = {9},
  publisher   = {Zenodo},
  version     = {1.1},
  doi         = {10.5281/zenodo.22958675},
  url         = {https://doi.org/10.5281/zenodo.22958675}
}

Also rendered at quantenergy.tech/spec/; the concept DOI 10.5281/zenodo.22958674 always resolves to the latest version. Dataset DOIs and this protocol DOI are separate and correspond through the version table inside the document.

3 · Main dataset v1.1.0 — the numbers on this site

Published 2026-04-19 · CC BY 4.0

Contents 360+ configurations, 0.5B–14B, RTX 4090D (Ada), RTX 5090 (Blackwell), A800 (Ampere), Tesla T4 (Turing); FP16/NF4/INT8/FP8; n = 2, CV < 2%

Cite it for anything on this site except the RTX 4090 column

@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}
}

4 · RTX 4090 (Ada) deep dive — the newest column

Published 2026-08-21 (measured 2026-07-24, 2026-08-19 and 2026-08-20) · CC BY 4.0

Contents one RTX 4090, five models 0.5B–7B × FP16/NF4/INT8, 15/15 real measurements from the open container; a second session of 10 measurements pairing energy with same-run perplexity (n = 1 each); and a third session repeating the five INT8 configurations three times each (n = 3, CV of the energy delta 0.6–3.9%)

Cite it for the RTX 4090 column, INT8-on-Ada and 7B-on-Ada claims

@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}
}

5 · llama.cpp GGUF F16 vs Q4_0 on RTX 4090 — the MLPerf Client comparison

Published 2026-09-04 (measured 2026-09-03) · CC BY 4.0

Contents 45 runs on one RTX 4090: three llama.cpp GGUF arms (F16, our Q4_0, and the Llama-3.1-8B-Instruct-Q4_0.gguf that MLPerf Client v2.0 ships) × three output lengths (64/320/576) × n = 5, randomized order, forced cooldown before every run, NVML hardware energy counter at 100 Hz, with the measurement scripts, environment snapshot and model SHA-256 hashes

Cite it for the llama.cpp F16-vs-Q4_0 energy rows and anything on the MLPerf Client page. Not an MLPerf result.

@dataset{zhang_ecocompute_llamacpp_gguf_2026,
  title     = {Independent replication of MLPerf-style energy measurement on
               NVIDIA RTX 4090 --- Llama-3.1-8B-Instruct (FP16 vs Q4\_0)},
  author    = {Zhang, Hongping},
  year      = {2026},
  publisher = {Zenodo},
  note      = {45 runs, n=5 per cell, three output lengths, randomized order,
               forced cooldown, NVML hardware energy counter; supersedes the
               provisional 18-run session published on quantenergy.tech},
  doi       = {10.5281/zenodo.22295184},
  url       = {https://doi.org/10.5281/zenodo.22295184}
}

The concept DOI 10.5281/zenodo.22295183 always resolves to the latest version of this record.

6 · RTX 5090 (Blackwell) four-precision session — the same card on a newer stack

Published 2026-09-20 (measured 2026-09-19/20) · CC BY 4.0

Contents one RTX 5090 (32 GB, 575 W, driver 580.76.05), five models 0.5B–7B. Container track: NF4 and INT8 against a per-session FP16 baseline, batch 1, 256 tokens × 10 iterations, run twice with a full instance restart between sessions (n = 2 per cell; |A − B| mean 2.1 pp, max 5.2 pp). Standalone track: torchao 0.18.0 FP8 weight-only and FP8 dynamic activation+weight, n = 3, generation-only window. Raw reports, both scripts and the environment snapshot included; four reports that fell back to interpolation are shipped labelled as such and are not measurements.

Cite it for RTX 5090 INT8, the FP8 rows, and the card–stack-pair evidence: on this same card the NF4 crossover sits near 4.8B on the v1.1.0 stack but near 1.8B on this newer stack, while the same newer stack left Ada at break-even — the crossover tracks the card–stack pair, not the stack alone. Not an MLPerf result, and not pooled with v1.1.0's rows for the same card.

@dataset{zhang_ecocompute_rtx5090_2026,
  title     = {EcoCompute: LLM Inference Energy on NVIDIA RTX 5090 (Blackwell)
               --- FP16/NF4/INT8/FP8, 0.5B--7B, direct NVML, with
               cross-restart repeatability},
  author    = {Zhang, Hongping},
  year      = {2026},
  publisher = {Zenodo},
  note      = {Supplementary single-platform session to dataset v1.1.0
               (10.5281/zenodo.19647290); bitsandbytes 0.50.2 NF4/INT8 repeated
               across a full instance restart (n=2, |A-B| max 5.2 pp) plus a
               separate torchao 0.18.0 FP8 track (n=3, generation-only window)},
  doi       = {10.5281/zenodo.22855133},
  url       = {https://doi.org/10.5281/zenodo.22855133}
}

The concept DOI 10.5281/zenodo.22855132 always resolves to the latest version of this record. The README inside the archive carries a 10.5281/zenodo.XXXXXXX placeholder in its BibTeX block, minted after the file was written — the block above is the correct one.