ecocompute-dynamic-eval/CITATION.cff at main · hongping-zh/ecocompute-dynamic-eval · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
cff-version: 1.2.0
message: "If you use this dataset or findings, please cite it as below."
title: "Energy Efficiency of Quantized Large Language Model Inference: Evidence for Quantization Efficiency Paradoxes on NVIDIA Blackwell, Ada Lovelace, and Ampere Architectures"
version: 1.0.0
date-released: 2026-02-14
authors:
  - family-names: Zhang
    given-names: Hongping
    email: zhanghongping1982@gmail.com
    affiliation: Independent Researcher
    city: Changsha
    region: Hunan
    country: CN
url: "https://github.com/hongping-zh/ecocompute-dynamic-eval"
repository-code: "https://github.com/hongping-zh/ecocompute-dynamic-eval"
repository-artifact: "https://github.com/hongping-zh/ecocompute-dynamic-eval/tree/main/metadata"
license: MIT
keywords:
  - large language models
  - quantization
  - energy efficiency
  - GPU inference
  - green AI
  - bitsandbytes
  - INT8 paradox
  - NF4 crossover
  - NVIDIA Blackwell
  - NVIDIA Ada Lovelace
  - NVIDIA Ampere
  - reproducible research
abstract: "We present empirical evidence of two quantization efficiency paradoxes in LLM inference across three NVIDIA GPU architectures. Through 32 rigorously-controlled measurements, we discovered that (1) default bitsandbytes INT8 increases energy by 17-147% due to mixed-precision decomposition overhead, and (2) NF4 exhibits a crossover threshold at approximately 5-6B parameters. Ablation experiments establish that disabling mixed-precision decomposition recovers +79-98% throughput and -35-41% energy. All data, metadata, and reproducibility artifacts are publicly available."
preferred-citation:
  type: article
  authors:
    - family-names: Zhang
      given-names: Hongping
      email: zhanghongping1982@gmail.com
  title: "Energy Efficiency of Quantized Large Language Model Inference: Evidence for Quantization Efficiency Paradoxes on NVIDIA Blackwell, Ada Lovelace, and Ampere Architectures"
  year: 2026
  journal: "arXiv preprint"
  notes: "32 measurements across RTX 5090 Blackwell, RTX 4090D Ada Lovelace, and A800 Ampere"