Assessing the environmental footprint of AI models at Meta

I was invited by Meta FAIR team to explain how to better account for AI environmental foorprint during the training phase. The objective of this talk was to provide pratical ways to enhance carbon emissions disclosure, such as what Meta already does for its Llama models.

I especially focused on how to broaden the scope of energy consumption accounting to include usage impacts. This involves not only GPU energy but also the entire AI server and data center Power Usage Effectiveness (PUE). I also introduced ways to account for embodied impacts (impacts related to hardware manufacturing) using the BoaviztAPI open-source tool.

Additionally, I listed relevant resources for AI researchers, including papers, tools, and best practices.