Sber and the AIRI Institute: Eco2AI A tool for assessing the carbon footprint of AI models

2022: Creating a Eco2AI Library

The laboratory artificial intelligence Sberbank , in partnership AIRI Institute of Artificial Intelligence with, created open source a library of ^[1] assess the equivalent carbon footprint , depending on the electricity spent on training models. Sberbank machine learning announced this on June 17, 2022, according to the Alexandra Vedyakhina First Deputy Chairman of the Management Board of Sberbank.

The library determines the available central and graphical computing devices, processors allows you to set regional norms for the carbon cost of generating electricity and calculates energy during code execution on devices.

Knowledge of the energy balance allows you to calculate how much each training session of a particular artificial intelligence model requires resources and what the equivalent carbon footprint leaves. The Eco2AI library allows you to track energy consumption and carbon footprint in the training process of artificial intelligence models, which creates an additional opportunity to optimize and reduce the cost of their learning processes.

Tracking the dynamics of energy consumption will help researchers assess the number of parameters necessary for the models to work, facilitate their architectures as much as possible and develop more optimal machine learning models in terms of computational complexity.

Sberbank is actively developing the Green AI initiative on the sustainable development agenda, which focuses on optimizing resources for deep learning. Our library helps track efficiency in terms of computational complexity. Of course, our development is in the public domain, since Sberbank is one of the drivers of ESG transformation countries and is interested in ensuring that other companies with huge computing power can use our methodology and reduce their carbon footprint, "said Alexander Vedyakhin, First Deputy Chairman of the Management Board of Sberbank.