Artificial intelligence news.
Straight from the original announcements.

NVIDIA cuBLAS adds FP64 algorithms for faster scientific simulations

November 13, 2026, 23:13By the AI News Bites Editorial TeamSource: NVIDIA Developer
NVIDIA cuBLAS adds FP64 algorithms for faster scientific simulations
Image: NVIDIA Developer

NVIDIA has added two algorithms to cuBLAS that accelerate double-precision matrix multiplication on integer Tensor Cores while preserving full FP64 accuracy. The company says CUDA Toolkit 13.4.2 delivers substantial gains for quantum chemistry and materials science workloads.

Key points
  • CUDA Toolkit 13.4.2 adds Ozaki-II and 2M algorithms to cuBLAS.
  • Ozaki-II accelerates real FP64 DGEMM on integer Tensor Cores.
  • The 2M method reduces complex multiplication from four real operations to two.
  • NVIDIA reports up to 6.6 times speedup in the cited application benchmarks.
  • Rubin GPUs support TI16 and tcgen05.mma acceleration for Ozaki computations.

Ozaki-II and 2M target high-precision workloads

Many scientific simulations spend much of their execution time on dense general matrix multiplication, known as GEMM. cuBLAS, NVIDIA's library for accelerating Basic Linear Algebra Subprograms on its GPUs, now uses the Ozaki-II and 2M algorithms to speed up real and complex double-precision calculations.

DGEMM handles real 64-bit floating-point numbers, while ZGEMM handles complex 64-bit floating-point numbers. NVIDIA says cuBLAS breaks double-precision values into integer pieces, processes the pieces on Tensor Cores and reassembles the result. The resulting calculations retain full double-precision accuracy.

Ozaki-II improves on the earlier Ozaki-I method by using the Chinese Remainder Theorem. It reduces the number of required integer matrix multiplications. The 2M method applies to complex matrix multiplication, reducing the operation from four real multiplications to two.

According to NVIDIA, cuBLAS can automatically accelerate DGEMM and ZGEMM without application code changes. Its Automatic Dynamic Precision framework selects an integer width at runtime, although users can also set it explicitly. If the requested accuracy cannot be maintained through the emulation path, cuBLAS falls back to native FP64 instructions.

Reported gains across NVIDIA systems

NVIDIA reports that DGEMM with Ozaki-II runs 5.5 times faster than native FP64 on NVIDIA VR NVL72 and 3.9 times faster on NVIDIA GB200. On NVIDIA GB200, the method is also 2 times faster than Ozaki-I.

For complex calculations, cuBLAS ZGEMM with the 2M algorithm achieves a 5 times speedup over native Complex FP64 on NVIDIA GB200, according to the company. NVIDIA says Ozaki-II DGEMM reaches up to 175, 147 and 37 TFLOPS on NVIDIA VR NVL72, NVIDIA GB200 and NVIDIA RTX PRO 6000, respectively.

The company also tested the algorithms in application scenarios. In a coupled cluster singles and doubles, or CCSD, quantum chemistry workload, Ozaki-II delivered end-to-end speedups over native FP64 of 2 times on NVIDIA VR NVL72 and 1.4 times on NVIDIA GB200. In a domain-specific mixed-precision mode, the reported gains were 5 times and 3.5 times, respectively.

In BerkeleyGW materials science simulations, NVIDIA reports that ZGEMM with the 2M algorithm achieved speedups of 6.6 times on NVIDIA VR NVL72 and 3.9 times on NVIDIA GB200 compared with native ZGEMM.

Rubin support and access through CUDA Toolkit 13.4.2

NVIDIA's Rubin GPUs introduce the TI16 numeric format, described by the company as a 12-bit integer in sign-magnitude format with a range of [-2047, +2047]. CUDA Toolkit 13.4 uses TI16 to accelerate Ozaki-I and Ozaki-II computations through the tcgen05.mma instruction on NVIDIA Rubin GPUs with compute capability 10.7.

The algorithms are included in CUDA Toolkit 13.4.2. NVIDIA directs developers to the floating point emulation section of the cuBLAS documentation for API details, including cublasSetEmulationStrategy and manual mantissa bit control. It also points users to cuBLAS Emulation samples on GitHub for evaluating the gains on their own workloads.

For researchers running quantum chemistry and materials science simulations, the change offers a way to use faster integer Tensor Core hardware for workloads that require double-precision results. The reported application benchmarks suggest that the benefit can extend beyond isolated matrix multiplication, although the results vary by GPU, algorithm and workload.

Related news

Dell expands AI Factory with storage, compute and governance tools
Chips and infrastructureNovember 16, 2026, 07:00

Dell expands AI Factory with storage, compute and governance tools

Dell Technologies is expanding its Dell AI Factory with NVIDIA with persistent storage, additional compute options and supported software aimed at moving agentic AI from…

NVIDIA Says cuPhoton Correlated a 10 TB EHT Dataset in 70 Seconds
Chips and infrastructureNovember 13, 2026, 22:17

NVIDIA Says cuPhoton Correlated a 10 TB EHT Dataset in 70 Seconds

NVIDIA says its cuPhoton.xBeam software processed a 10 TB radio astronomy dataset in 70 seconds on a system with 64 GPUs. The result, according to NVIDIA, was about 460…

NVIDIA outlines five ways to make every AI memory byte work harder
Chips and infrastructureNovember 5, 2026, 17:00

NVIDIA outlines five ways to make every AI memory byte work harder

NVIDIA has outlined five co-designed technologies aimed at increasing the amount of useful AI work produced by each byte of memory. The company says the approach enables…