NVIDIA cuBLAS adds FP64 emulation algorithms for faster simulations

NVIDIA added two double-precision algorithms to cuBLAS in CUDA Toolkit 13.4.2, aimed at scientific simulations that spend most of their runtime on dense matrix multiplication. Ozaki-II uses the Chinese Remainder Theorem to require fewer integer multiplications, and the 2M method cuts a complex matrix multiplication from four real multiplications to two. Both run on integer Tensor Cores while preserving full FP64 accuracy, and cuBLAS falls back to native FP64 when the precision a run needs exceeds the configured limit. NVIDIA reports 5.5x and 3.9x DGEMM speedups over native FP64 on VR NVL72 and GB200. Applications get the acceleration without code changes; the algorithms ship in the toolkit, with API controls and GitHub samples available.