Oz-FP4: Emulating FP64 DGEMM on Low-Precision FP4 Tensor Cores

teortaxesTex · x · 2026-08-11

A new paper introduces Oz-FP4, a method to emulate FP64 DGEMM (double-precision matrix multiplication) using low-precision FP4 Tensor Cores.

By constructing Ozaki schemes I and II, the approach enables high-precision matrix multiplication on low-precision arithmetic units, offering a new way to maximize the utility of modern AI hardware.

Original post →

More from Infra

Infra channel →