Nvidia adds cuBLAS algorithms to accelerate double-precision simulations

Nvidia says cuBLAS in CUDA Toolkit 13.4.2 accelerates scientific matrix calculations by moving double-precision arithmetic onto integer Tensor Cores while preserving full double-precision accuracy. The library applies the acceleration without application code changes. The Ozaki-II algorithm reduces the integer multiplications needed to reconstruct accurate results, while the 2M method reduces complex matrix multiplication from four real matrix multiplications to two. Nvidia reports a 3.9x end-to-end speedup for the BerkeleyGW materials-science application on GB200. Gains depend on hardware and matrix size. When the required precision exceeds the configured limit, cuBLAS falls back to native double-precision instructions.