logo Frugale

Tool

EcoLogits

Edited by :
GenAI Impact

Summary

EcoLogits monitors energy consumption and environmental impact from the use of generative AI models via APIs (environmental impacts and energy consumption of LLMs during inference).

The goal is twofold: to provide users with transparent information about the environmental cost of their AI usage and to help developers better account for (and reduce) the environmental impact of their projects.

EcoLogits relies on a precise and detailed methodology, including life cycle analysis and multi-criteria evaluation. It assesses energy consumption, global warming potential (GWP) expressed in CO₂ equivalents, abiotic resource consumption, and primary energy usage.

At present, EcoLogits focuses on inference impact, meaning the phase where generative AI models process user requests.

EcoLogits is a tool designed to be deployed by technical development teams within their projects. It is not intended for direct public use.

The editorial perspective

EcoLogits, developed by the non-profit organization GenAI Impact, appears to be a highly useful tool. It was created during a hackathon organized by Data For Good and benefited from input and support from a large community.

Notably, Boavizta, a well-known association specializing in evaluating the environmental footprint of digital technologies, contributed to the project. The methodology is explained with great precision and pedagogy, making it accessible even to those who are less familiar with these topics.

The tutorial section is designed to be highly accessible to any technical development team—and even to anyone with basic command-line and programming skills.

EcoLogits heavily relies on data from Hugging Face, which hosts numerous generative AI models. Hugging Face itself measures the footprint of the models it hosts using CodeCarbon, a tool incubated by Data For Good.

To provide carbon footprint estimates for a wide range of AI models, including proprietary models (such as those from Mistral AI and OpenAI), GenAI Impact chose to extrapolate data based on Hugging Face’s dataset.

In brief, the editorial perspective

The most

  • Open-source tool
  • Detailed methodology
  • A very well-detailed tutorial to facilitate adoption by development teams
  • Tool developed in collaboration with expert organizations

The least

  • The documentation could more directly explain the target audience and highlight use cases. Simply making the link with the Ecologits Calculator tool more prominent would suffice
  • The impact calculation of closed-source models and/or those not hosted on Hugging Face is based on extrapolation, which implies a degree of inaccuracy
  • Attention, the resource is only available in English

Publication date

June 2024

Available in

  • English

License

All methodologies are licensed under CC BY-SA 4.0