As modern computing systems become increasingly constrained by energy costs, data movement, and especially memory bandwidth, mixed-precision computing is emerging as a key strategy for improving scientific application performance while maintaining acceptable numerical accuracy. This topic is particularly timely in HPC, where low-precision hardware capabilities are advancing rapidly, but many scientific workloads remain limited by accuracy and stability requirements and bandwidth-bound execution.
This session, targeted at HPC researchers and application developers, brings together two complementary perspectives on how precision can be treated as a tunable design parameter in scientific computing. The first talk focuses on how mixed-precision methods can be engineered to breach the memory wall in bandwidth-limited scientific applications. The second examines the challenges in automated floating-point precision tuning for scientific software. Together, the talks present practical techniques for overcoming current hardware and algorithmic bottlenecks. Attendees will gain a clear view of why mixed precision matters now, what technical challenges remain, and how current methods can improve efficiency, scalability, and performance in scientific workloads.
