EUVIMEDEUVIMED entdecken
Crossref

Green AI benchmarking: A framework for accuracy, energy, and carbon tradeoff evaluation

Khawaja Moinuddin Jamal, Muhammad Nadeem, Danish Jamil

Perspectives in Social Sciences · 30.10.2026 · 5 (10(II)) · 1544-1569

DOI: 10.63468/pss.178 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Perspectives in Social Sciences, online erschienen am 30.10.2026. Verfasst von Khawaja Moinuddin Jamal und 2 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

This study proposes and evaluates a Green AI Benchmarking framework for comparing artificial intelligence models through the joint assessment of task accuracy, operational energy consumption, and carbon emissions. Responding to the limitations of accuracy-only leaderboards, the framework treats each model-task-run as an auditable record containing quality metrics, latency, throughput, power traces, facility-adjusted energy, carbon-intensity metadata, and a sustainability-adjusted decision score. The empirical evaluation used 48 controlled benchmark observations across eight model configurations, including image-classification, text-classification, and question-answering workloads, with six repeated runs per model. Results show that higher accuracy does not automatically justify higher environmental cost. DistilBERT-base achieved the strongest overall Green AI Score because it combined high text-classification performance with the lowest energy and carbon values, while MobileNetV3-Small emerged as the most sustainable image-classification candidate despite not having the highest vision accuracy. Larger models such as RoBERTa-base and ViT-B/16 delivered marginal accuracy gains but required substantially greater energy and carbon expenditure. The findings demonstrate that model selection should begin with a minimum quality threshold, then evaluate energy, carbon, and Pareto trade-offs. The framework offers a reproducible, decision-oriented approach for sustainable AI reporting, procurement, deployment, and governance. It supports transparent communication between researchers, engineers, managers, and policy stakeholders in practice.

Zitieren

APA

Jamal, K. M., Nadeem, M., & Jamil, D. (2026). Green AI benchmarking: A framework for accuracy, energy, and carbon tradeoff evaluation. Perspectives in Social Sciences, 5(10(II)), 1544-1569. https://doi.org/10.63468/pss.178

Vancouver

Jamal KM, Nadeem M, Jamil D. Green AI benchmarking: A framework for accuracy, energy, and carbon tradeoff evaluation. Perspectives in Social Sciences. 2026;5(10(II)):1544-1569. doi:10.63468/pss.178

BibTeX
@article{Jamal2026,
  author = {Jamal, Khawaja Moinuddin and Nadeem, Muhammad and Jamil, Danish},
  title = {Green AI benchmarking: A framework for accuracy, energy, and carbon tradeoff evaluation},
  journal = {Perspectives in Social Sciences},
  year = {2026},
  volume = {5},
  number = {10(II)},
  pages = {1544--1569},
  doi = {10.63468/pss.178}
}
Kostenlos starten →