Open-source tool tracks AI computing's CO2 footprint

December 03, 2020 //By Rich Pell
Open-source tool tracks AI computing's CO2 footprint
AI experts at deep learning research institute Mila, data science and advanced analytics group BCG GAMMA, Haverford College (Philadelphia, PA), and machine learning platform have jointly created an open-source software package designed to estimate the location-dependent CO2 footprint of computing.

The software, called CodeCarbon, is a lightweight software package that seamlessly integrates into Python codebase. It estimates the amount of carbon dioxide (CO 2) produced by the computing resources used to execute the code to incentivize developers to optimize their code efficiency.

It also advises developers on how they can reduce emissions by selecting their cloud infrastructure in regions that use lower carbon energy sources, say the software's developers.

"AI is a powerful technology and a force for good, but it's important to be conscious of its growing environmental impact," says Yoshua Bengio, Mila founder and Turing Prize recipient. "The CodeCarbon project aims to do just that, and I hope that it will inspire the AI community to calculate, disclose, and reduce its carbon footprint."

Sylvain Duranton, a managing director and senior partner at Boston Consulting Group (BCG) and global head of BCG GAMMA, adds, "If recent history is any indicator, the use of computing in general, and AI computing in particular, will continue to expand exponentially around the world. As this happens, CodeCarbon can help organizations make sure their collective carbon footprint increases as little as possible."

The software's developers say such a tool is needed to help address the amount of energy needed to support the massive computing behind AI. For example, training a powerful machine-learning algorithm can require running multiple computing machines for days or weeks.

And, say the developers, the fine-tuning required to improve an algorithm by searching through different parameters can be especially intensive. For recent state-of-the-art architectures like VGG, BERT, and GPT-3, which have millions of parameters and are trained on multiple GPUs (graphic processing units) for several weeks, this can mean a difference of hundreds of kilograms of CO 2eq - a standardized measure used to express the global warming potential of various greenhouse gases.

The software's emission tracker records the amount of power being used by the underlying infrastructure

Vous êtes certain ?

Si vous désactivez les cookies, vous ne pouvez plus naviguer sur le site.

Vous allez être rediriger vers Google.