Faster Scientific Simulations with New FP64 Algorithms in NVIDIA cuBLAS
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
NVIDIA's cuBLAS in CUDA Toolkit 13.4.2 now emulates FP64 matrix multiplication on integer Tensor Cores, using the Ozaki-II scheme for DGEMM and a 2M method that halves the real GEMMs needed for ZGEMM. NVIDIA reports DGEMM gains of 3.9x on GB200 and 5.5x on VR NVL72, and BerkeleyGW speedups of 3.9x and 6.6x. These figures come from NVIDIA's own benchmarks, and the accuracy match across BerkeleyGW configurations couldn't be independently verified. Applications need no code changes, and cuBLAS falls back to native FP64 when accuracy requires it.