SignalTech/AI
Faster Scientific Simulations with New FP64 Algorithms in NVIDIA cuBLAS
New FP64 algorithms in NVIDIA cuBLAS enable faster scientific simulations.
Context
What happened: NVIDIA introduced new cuBLAS algorithms, Ozaki-II and the 2M method, in CUDA Toolkit 13.4.2 that run FP64 DGEMM and ZGEMM on integer Tensor Cores while keeping full double-precision accuracy, with no application code changes.
Key numbers: DGEMM with Ozaki-II runs 5.5x and 3.9x faster than native FP64 on NVIDIA VR NVL72 and NVIDIA GB200, and 2x faster than Ozaki-I on
Why it matters: cuBLAS no longer has to choose between accuracy and speed, and it falls back to native FP64 MMA instructions if accuracy targets are not met, so applications can use it in production deployments.
What changed
NVIDIA introduced new cuBLAS algorithms, Ozaki-II and the 2M method, in CUDA Toolkit 13.4.2 that run FP64 DGEMM and ZGEMM on integer Tensor Cores while keeping full double-precision accuracy, with no application code changes.
Why it matters
cuBLAS no longer has to choose between accuracy and speed, and it falls back to native FP64 MMA instructions if accuracy targets are not met, so applications can use it in production deployments.