Systems-Facing Track – April 2026

Floating Point Precision in Modern GPU Infrastructure: Practical Considerations for HPC Architecture and Procurement

With the emergence of new GPU architectures, those building or refreshing HPC systems are facing additional decision-making around platform selection. Previous assumptions that the latest hardware will always be the best choice are being challenged by floating point precision, with some newer GPU models performing more poorly than older ones for certain workloads. This talk will offer a practical overview of precision across modern accelerator environments and how these factors affect both system architecture and procurement strategy. This presentation will draw on Georgia Tech’s experience both experimenting with emerging architectures and deploying reliable, highly scalable HPC systems for its research community.

Speakers: Aaron Jezghani, PhD, Associate Director, Platform & Architecture and Scheduler Architect, PACE, Georgia Tech; Will Powell, Research Technologist, College of Computing, Georgia Tech