AI Energy Score

Overview

AI Energy Score is an initiative to establish standardized energy efficiency ratings for AI model inference. It evaluates models on 10 common tasks and gives 1-5 star ratings within each task or model-size class, based on measured GPU energy in watt-hours. Results are published on a public leaderboard, and developers submit measurements through a submission portal hosted on Hugging Face. The project focuses on inference energy rather than training.

Use cases

• Compare energy efficiency of AI models
• Benchmark models for procurement
• Track inference energy by task
• Submit model energy measurements

Sources

• AI Energy Score
• AIEnergyScore MIT licence