AI Energy Score v2: Refreshed Leaderboard, Now with Reasoning
5 days ago
- #AI Energy Efficiency
- #Sustainable AI
- #Benchmarking
- Launch of AI Energy Score v2 leaderboard with new models and reasoning task benchmarking.
- Improved benchmarking code and submission process for streamlined evaluation.
- Reasoning models use 30 times more energy on average than non-reasoning models.
- Energy increase due to reasoning ranges from 150 to 700 times more for specific models.
- Reasoning models generate 300-800 times more tokens, leading to higher energy use.
- Energy use of reasoning models is less predictable compared to standard LLMs.
- Newer models show mixed efficiency results, with some using more energy than older models.
- Salesforce integrates AI Energy Score into internal benchmarking, promoting energy transparency.
- Future plans include adding video generation and agentic tasks to the benchmark.
- Community support is crucial for the future of AI Energy Score.