Abstract

The number of Artificial Intelligence applications being developed and deployed is continually increasing. The effects of these activities on the biosphere, particularly on climate change, have attracted attention since 2019, but assessment methodologies still require improvement. More advanced evaluation methods and a deeper understanding of these impacts are necessary to minimize the environmental impacts of artificial intelligence.

With an emphasis on the training phase, this thesis investigates how machine learning (ML) affects the environment.

First, we question how the electricity consumption of IT infrastructures is measured by comparing power meters currently in use with different benchmarks and infrastructures, focusing on Graphic Processing Units (GPUs). These findings are used to analyze the electricity required to train models selected from the MLPerf benchmark on various ML infrastructures, ranging from an edge device to a supercomputer. Finally, the thesis shifts toward examining the more general environmental impacts of ML, based on an estimation of the embodied impacts of ML infrastructures. These impacts are allocated to each model training, enabling a comparison with the impacts of electricity usage. While numerous ML environmental impact indicators exist, this study focuses on primary energy consumption, global warming potential, and abiotic depletion potential for minerals and metals.

In conclusion, this thesis proposes a methodology that enables a reproducible multi-criteria evaluation of the impact of machine learning training on the environment and can be applied to different ML infrastructures, thus enabling fair comparison and enlightened choices.