Workshop: Sustainable Supercomputing
Authors: Varsha Rao and Clarisse Cheung (University of Chicago) and Andrew A. Chien (University of Chicago, Argonne National Laboratory)
Abstract: The carbon footprint of computing systems is growing rapidly, driven by exponential AI growth. Many studies have suggested approaches such as fast hardware upgrades, using renewable power, and extending system lifetime to improve overall carbon efficiency and reduce carbon footprint. We analyzed carbon efficiency of the top 500 high performance computing (HPC) systems. It shows that a combination of key factors affects carbon efficiency. We propose tempered hardware upgrade, where hardware upgrades are performed while considering carbon optimization, accounting for both new and old hardware's embodied carbon. We evaluate several approaches and find that optimizing for just one factor: performance, operational carbon, or embodied carbon fails to improve carbon efficiency, sometimes worsening it or yielding negligible gains. For example, a fast hardware upgrade worsens embodied carbon efficiency by 0.9x but slightly improves operational carbon efficiency by 1.1x, however increasing embodied and operational carbon emissions by 125.5% and 47.2%, respectively. In contrast, tempered hardware upgrade improves embodied carbon efficiency by 1.7x and operational carbon efficiency by 19x, while increasing embodied carbon emissions by only 12.7% and reducing operational carbon emissions by 86.2%. We conclude that a balanced approach is required to improve carbon efficiency and reduce carbon footprint of systems.
Back to Sustainable Supercomputing Archive Listing
Back to Full Workshop Archive Listing