Machine Learning (ML) workloads have rapidly grown in importance, but raised concerns about their carbon footprint; previous studies have produced very large estimates for that cost. Based on an in-depth study of current and past ML energy consumption, we describe four best practices to reduce ML training energy by up to 100x and CO2 emissions up to 1000x, show that the cost and carbon footprint of ML training is 100x–100,000x lower than previously reported, and explain the sources of these discrepancies. If the whole ML field adopts these best practices, we predict that total gross carbon emissions from training will plateau and that net carbon emissions from ML training will approach zero.

Download the slides for this talk.Download ( PDF, 7353.04 MB)