Power Sloshing in Compound Servers for Large-Scale AI Inference Workloads
Published in 53rd IEEE/ACM International Symposium on Computer Architecture (ISCA 2026), 2026
Albert Cho, Jovan Stojkovic, Leonardo Piga, Abhishek Dhanotia, Sultan Mahmud Sajal, Gefei Zuo, Krishna Malladi, Devon Akers, Kalyan Subramanian, Shobhit Kanaujia, Alexandros Daglis
Identified power management for AI inference in compound servers as a rising problem, share our observations and insights on the matter based on Meta’s fleetwide data, and suggest a simple scheme for quick deployment at scale that redistributed power across components based on the current bottleneck.