Environmental Impacts of GenAI with EcoLogits at ADIRA

In this talk for the ADIRA, we explore the environmental impacts of generative AI. We highlight the difference between training and inference impacts, as well as the growing environmental footprint of new LLMs. We also emphasize the need for a reporting standard to better assess environmental impacts, including hardware manufacturing and multiple criteria (carbon footprint, abiotic resources, water consumption, etc.).

We discuss the beginnings of our project and how we created a methodology to assess the environmental impacts of LLMs at inference, based on open Life Cycle Assessment studies and energy benchmarks. We introduce our open-source tool, EcoLogits, implemented in Python, which assesses the environmental footprint of GenAI services at each request, supporting OpenAI, Anthropic, Mistral AI, and many others.

Finally, we provide best practices on how to reduce the environmental footprint of AI and IT in organizations. These include practicing frugal AI, questioning the necessity of AI and GenAI implementations, and offering tips on how to reduce the number of new devices in an organization.