Faster Scientific Simulations with New FP64 Algorithms in NVIDIA cuBLAS
Many scientific simulations, from molecular chemistry to materials science, spend the majority of their execution time computing high-precision dense general...
First seen by Oriel on .
AI summary
This NVIDIA post describes cuBLAS algorithms that use integer Tensor Cores to accelerate full-accuracy FP64 DGEMM and ZGEMM, reporting substantial speedups in quantum chemistry, materials science, and matrix benchmarks.
- CUDA Toolkit 13.4.2 automatically applies Ozaki-II and 2M to DGEMM and ZGEMM without application code changes, falling back to native FP64 when necessary.
- Ozaki-II uses Chinese Remainder Theorem reconstruction, reducing GEMM growth from quadratic to linear versus Ozaki-I while tiled execution limits workspace usage.
- The 2M method computes complex matrix products with two real GEMMs instead of four, enabled by exact integer operations that preserve accuracy.
- DGEMM with Ozaki-II reaches 5.5x native FP64 performance on NVIDIA VR NVL72 and 3.9x on NVIDIA GB200, while RTX PRO 6000 reaches 23x.
- CCSD and BerkeleyGW retain reported FP64-equivalent accuracy while achieving application speedups of up to 5x and 6.6x respectively with mixed or emulated precision.
#cublas #fp64 #tensor cores #scientific computing
Generated by gpt-5.6-sol. Check the original for details.
Aliases
| Source | External ID | URL |
|---|---|---|
| NVIDIA Developer Blog | https://developer.nvidia.com/blog/?p=125124 | https://developer.nvidia.com/blog/faster-scientific-simulations-with-new-fp64-algorithms-in-nvidia-cublas/ Read discussion |