Breaking02 / IT
NVIDIA adds faster double-precision matrix calculations
NVIDIA says new cuBLAS algorithms accelerate matrix multiplication on integer Tensor Cores while preserving double-precision accuracy.

NVIDIA said it introduced new matrix calculation acceleration in cuBLAS, its GPU maths library, in CUDA Toolkit 13.4.2. The company says the approach preserves double-precision accuracy.
According to NVIDIA, cuBLAS splits double-precision (FP64) numbers into integer pieces. It multiplies them on integer Tensor Cores, specialised GPU calculation units, then reconstructs the result. The method applies to both real and complex matrix multiplication.
NVIDIA says Ozaki-II requires fewer integer multiplications than the earlier Ozaki-I method. Its 2M algorithm reduces the real matrix multiplications needed for a complex matrix product. It cuts the number from four to two.
In NVIDIA's GB200 measurements, real matrix multiplication ran up to 3.9 times as fast as native FP64. The company also compared coupled cluster singles and doubles (CCSD), a quantum chemistry calculation method. It reported a 1.4-fold speedup for the complete CCSD calculation on GB200 while retaining FP64 precision.
NVIDIA's technical post describes cases where cuBLAS selected native FP64 for small matrices. The company says this followed an automatic assessment that acceleration offered no advantage. The library also falls back to native FP64 when the required precision exceeds a configured limit.
NVIDIA has not disclosed the GPU power settings used in these performance comparisons.