Towards LLM Energy Labels: Accuracy and Energy Efficiency
Levi Ari Pronk, Ocean Wang, Nicholas Wu, Madhav Chawla, Yasar Saltuk Bugra Kocdas.
Group 26.
Paper.Source code.We propose Energy Per Correct Answer (EPCA), a metric combining energy cost and accuracy, and a proof-of-concept tool that benchmarks LLMs across coding, math, and logical reasoning domains.
We propose Energy Per Correct Answer (EPCA), a composite metric that jointly captures inference energy cost and task accuracy for LLMs. Inspired by the EU energy labelling framework for appliances, we built a proof-of-concept tool that benchmarks locally hosted LLMs by measuring GPU energy consumption via NVIDIA power telemetry, evaluating correctness, and assigning an A-G energy label. We tested our tool across three domains: LeetCode-style coding problems (pass@1), mathematical reasoning, and logical reasoning (exact-match accuracy). Our tool, including a web-based leaderboard, is publicly available to encourage transparent reporting of LLM energy efficiency.