Assessing the ecological impact of LLMs: the innovative Ecologits initiative

The rise of Generative Artificial Intelligence (GenIA) and Large Language Models (LLM) technologies has revolutionized many sectors, from content creation to the provision of personalized services. Today, Generative Artificial Intelligence is increasingly interwoven into our lives through tools (ChatGPT, Midjourney, DALL-E, etc.) capable of generating content (text, visual and video) of near-human quality and at unrivalled speed. However, this technological advance is accompanied by considerable energy consumption and a significant environmental impact. It is therefore crucial toassess these impacts to promote a reasoned and responsible use of generative AI and LLMs in particular. In this article, we explore why it is necessary to measure the ecological impact of generative AI and how the Ecologits.ai project contributes to this process with an innovative and impactful approach.

Performance and cost: an incomplete assessment of LLMs

The evaluation of LLMs is based on several key criteria, including technical performance, contextual relevance, ethics, and security, as well as considerations of efficiency and cost. Today, the performance and cost of language models are often the primary concerns of the companies that use them. However, this evaluation remains incomplete if the environmental impact of these technologies is not taken into account.

Indeed, LLMs require huge amounts of energy for training and inference, and their increasing use contributes significantly to greenhouse gas emissions. According to some studies, training large models could generate as much CO2 as a car over its lifetime. In view of this, measuring only the performance and cost of LLMs is tantamount to ignoring a crucial part of their overall impact. However, while the performance and cost of generative AI are easy to assess, the environmental impact of its use remains a considerable grey area, despite - or perhaps because of - the incredible speed of its adoption by the general public.

In this context, the organization Data for Good, which works each year on projects that have a positive impact on social, societal, and environmental issues, has conducted in-depth research on the challenges of generative AI and its direct and indirect implications for our society and environment. About a hundred volunteers have formed the “GenAI Impact” collective to highlight the environmental impact of GenAI technologies, better understand and quantify them, and raise awareness among users. The goal? To better assess the environmental impact of generative AI and enable its thoughtful and responsible use.

That is why the “GenAI Impact” project team created 𝐄𝐜𝐨𝐋𝐨𝐠𝐢𝐭𝐬, a Python library designed to enable users to measure the environmental impacts of queries made to various generative AI applications.

Ecologits.ai: a pioneering initiative

The first version of Ecologits.ai is an innovative Python library that evaluates the environmental impact and energy consumption of LLMs during inference. This module is in line with a logic of transparency and environmental responsibility, providing developers and companies with the tools they need to quantify and reduce the ecological footprint of their models.

Ecologits.ai features

1. Measuring energy consumption: Ecologits.ai provides tools for measuring the amount of energy consumed during model inference. This includes granularity to see energy consumed per query, per session, or over longer periods.
2. Carbon footprint calculation: Based on energy consumption data and using specific conversion factors, the library calculates the carbon footprint generated by LLM use.

A transparent, collaborative approach

One of Ecologits.ai's strengths is its transparent, collaborative approach. The project is open-source, allowing the community to contribute, improve and verify the algorithms used. This transparency guarantees that the measurements taken are reliable and accepted by the community.

The aim of this tool is to foster a sustainable approach within the technology industry, where developers can make informed choices to minimize environmental impact without sacrificing performance.

In concrete terms, EcoLogits tracks the energy consumption and environmental footprint of AI models used for inference. Models belonging to five Artificial Intelligence services (OpenAI, MistralAI, Hugging Face, Cohere, Anthropic, Google) can already be analyzed.

Wikit's contribution to Ecologits

As specialists in creating chatbots based on generative AI and committed to responsible digital practices, we felt it was important to contribute to a project like Ecologits. Our Research & Development team identified a key area for optimizing the library: promoting the use of Ecologits in production by offering integration with the LiteLLM gateway. This open-source project facilitates the invocation of large language models (acting as a sort of proxy) by standardizing calls to various providers in the OpenAI format.

Why is this necessary? Having to individualize call parameters for each supplier greatly complicates the architecture of our applications. Unifying them in a single format makes the application more flexible and facilitates the integration of new models and/or suppliers, while keeping the code simple.

We have therefore contributed to Ecologits by proposing integration with LiteLLM. Our objective is clear: to facilitate the integration of environmental impact measurement in production. We are still actively proposing improvements to the project so that it becomes a mandatory best practice in the creation of architectures based on large language models.

Conclusion

Thanks to solutions like Ecologits.ai, companies now have an invaluable tool for accurately assessing the ecological impact of LLMs. This enables them to choose models that combine performance and sustainability. By adopting a holistic and innovative approach, it is possible to minimize the trade-off between ecology and performance, by making technological choices that support both business and environmental objectives. Performance and cost must no longer be the sole criteria for evaluating LLMs; it's time to integrate ecological impact into this equation for a more sustainable future!