cuBLAS uses integer Tensor Cores for FP64 matrix products
NVIDIA describes new cuBLAS algorithms that compute double-precision matrix products using faster integer Tensor Cores. Ozaki-II reconstructs products from modular integer calculations, reducing the multiplication work required by an earlier method. The library targets FP64-equivalent accuracy and falls back to native double-precision instructions when the needed precision exceeds configured limits. Reported gains depend on the hardware and numerical workload.
NVIDIA Technical Blog · Nov 13, 2026
Why it matters
Faster low-precision hardware can help scientific computing when the algorithm preserves the required numerical accuracy.
More context
The algorithms are available in CUDA Toolkit 13.4.2. Review cuBLAS floating-point emulation controls and try its samples on a representative matrix workload. Keep full-precision and deliberately reduced-precision modes distinct when comparing application results.
Keep in mind
Vendor algorithms and benchmarks on supported NVIDIA hardware. Exact integer intermediates do not mean every floating-point application becomes mathematically exact or equally faster; validate scientific outputs and end-to-end runtime.
Original read