Will FP64 Emulation Replace Native FP64 Computing?

Industrial Problems Seminar

Kasia Swirydowicz
AMD

Abstract

Ozaki Scheme I, Ozaki Scheme II, the Henry scheme, and similar methods are used to emulate GEMMs using lower-precision arithmetic. These methods typically use FP8 or INT8, although FP4 methods have also been proposed. Their popularity is growing in the HPC community as hardware evolves in response to advances in AI. With each generation, GPUs deliver more hardware-accelerated, low-precision matrix cores, while FP64 matrix performance remains flat or even declines. In my talk, I will explain how these methods work and discuss their benefits, such as reproducibility; their limitations, including memory requirements; and how they fit into the broader HPC computing landscape.