NVIDIA cuBLAS Achieves Up to 5.5x DGEMM Acceleration Using Integer Tensor Cores While Maintaining FP64 Precision
·2026.11.19 11:30
Key point
CUDA 13.4 achieves up to 5.5x DGEMM acceleration while maintaining FP64 precision using the Ozaki-II and 2M algorithms.
Details
- NVIDIA updated its FP64 emulation feature utilizing integer Tensor Cores in CUDA Toolkit 13.4.
- It recorded up to 5.5x DGEMM acceleration on Vera Rubin, 3.9x on GB200, and up to 23x on the RTX PRO 6000 Blackwell Server Edition.
- It applied the Ozaki-II algorithm based on the Chinese Remainder Theorem (CRT) and the 2M algorithm, which computes complex matrix multiplication as two real GEMMs.
- Automatic Dynamic Precision (ADP) determines the required number of guard bits and falls back to native FP64 instructions when limits are exceeded to ensure precision.
- In CCSD quantum chemistry simulations, it maintained energy errors below 10^-14 Eh with FP64 precision settings and showed a 4.7x speedup on Vera Rubin compared to A100.
- Activation requires setting the CUBLAS_EMULATE_DOUBLE_PRECISION=1 environment variable, and CUDA 13.4 or later and GPUs from Ampere onward are required.
- The CUDA 13.5 release notes mention an Ozaki-II complex operation error (CUB-10730) and a fixed-point emulation error (CUB-10744), requiring caution when using that version.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.