Workshops Archive

Skip the SuperPOD - The User-Driven Case for Throughput-Optimized GPU Clusters


Workshop: HUST-26: 13th International Workshop on HPC User Support Tools

Authors: Preston Smith and Ryan DeRue (Purdue University), Scott Michael and Kanin Bender (Indiana University), Greg Bauer (University of Illinois Urbana-Champaign), Christopher Heller (University of Illinois at Urbana-Champaign), Kevin Hildebrand (University of Maryland), Brock Palen (University of Michigan), Brian O'Shea (Michigan State University), Derek Weitzel and Caughlin Bohn (University of Nebraska-Lincoln), Janna Nugent (Northwestern University), Doug Johnson and Samuel Khuvis (Ohio Supercomputer Center), Dan Majchrzak (University of Oregon), Vlad Kholodovych and Exequiel Punzalan (Rutgers University), and Xiao Zhu (University of Washington)

Abstract: AI compute architectures are optimized for theneeds of the market, which today is hyperscaler AI training. It is critical for academic cyberinfrastructure leaders to understand the workload profiles in higher education and compare these to peer institutions to account for any local factors. This study seeks to compare and contrast the workload profiles across a set of university HPC centers to identify whether usage patterns are consistent across institutions and build out a data-driven model to capacity plan for GPU compute demand.


Back to HUST-26: 13th International Workshop on HPC User Support Tools Archive Listing

Back to Full Workshop Archive Listing