NVIDIA · FP64 GEMM Emulation

cuBLAS Ozaki-II and 2M FP64 Algorithms

NVIDIA's CUDA Toolkit 13.4.2 adds Ozaki-II and 2M algorithms to cuBLAS, which run double-precision (FP64) matrix multiplication on integer Tensor Cores while keeping full FP64 accuracy. The page reports DGEMM with Ozaki-II running 5.5x faster than native FP64 on VR NVL72 and 3.9x on GB200, and ZGEMM with the 2M method reaching 5x over native complex FP64 on GB200. Application results include a CCSD quantum chemistry case at 2x and 1.4x end-to-end speedups in full FP64, and BerkeleyGW at 6.6x and 3.9x. The change needs no application code changes and is available through the CUDA Toolkit.

Open the source ↗

Similar Developer Tools releases

More from NVIDIA

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.