An AI Energy Score, a benchmarking tool that lets AI developers and users evaluate the energy consumption of AI models, has been released by Salesforce, in collaboration with Hugging Face, Cohere, and Carnegie Mellon University
Salesforce announced it will be the first AI model developer to disclose the energy efficiency data of its proprietary models under the new framework.
The AI Energy Score aims to address the lack of transparency about the environmental impact of AI models. Similar to how Energy Star transformed energy efficiency standards for appliances and electronics, this initiative aims to establish a clear, trusted benchmark for AI model sustainability.
“Reducing AI energy consumption lowers operational costs, optimises infrastructure, and enhances long-term sustainability and profitability,” says Linda Saunders, Salesforce country manager and senior director of solution engineering for Africa.
The AI Energy Score debuted this week at the AI Action Summit in Paris, where leaders from over 100 countries, the private sector, and civil society convened, ostensibly to harness AI for good. By enhancing transparency, the score can drive market preference for efficient models and incentivise sustainable AI development. Recognised by the French Government and the Paris Peace Forum, the AI Energy Score features:
- Standardised Energy Ratings: A standardised framework for measuring and comparing AI model energy efficiency.
- Public Leaderboard: A comprehensive leaderboard that features scores for 10 common AI tasks — such as text generation, image generation, and summarisation — performed by 166 models, including Salesforce’s SFR-Embedding, xLAM, and SF-TextBase.
- Benchmarking Portal: A platform where AI developers can submit their open or proprietary AI models to be evaluated and added to the leaderboard. Open models can be automatically tested, while closed models can be evaluated through a secured testing sandbox.
- Recognisable Energy Use Label: A new 1- to 5-star label that rates AI model energy use, with five stars indicating the highest efficiency. This helps developers and users easily identify and choose more sustainable models. Once rated, AI developers can generate standardised labels to share their models’ energy score, with built-in guidance on the proper label display for visibility and impact.