AI · Oct 28, 2026
RTX Spark laptops get open speech models and a dial for background agentscuBLAS now fakes double precision on integer cores, and scientists get up to 6.6x speedups
· Updated · Signal Desk newsroom
Analysis of reporting by NVIDIA Developer. How we work

CUDA 13.4.2 brings Ozaki-II and 2M algorithms that deliver FP64-accurate matrix math on integer Tensor Cores. It’s Nvidia’s answer to fading native FP64, and it works, with a fallback when accuracy demands it.
Key takeaways
- cuBLAS in CUDA Toolkit 13.4.2 adds Ozaki-II for DGEMM and a 2M method for complex ZGEMM, running FP64-accurate math on integer Tensor Cores.
- DGEMM with Ozaki-II runs 5.5x faster than native FP64 on Vera Rubin NVL72 and 3.9x on GB200, and 2x faster than Ozaki-I on GB200.
- Real applications gain too: CCSD quantum chemistry runs 2x faster on VR NVL72 in full FP64, and BerkeleyGW runs 6.6x faster.
- On RTX PRO 6000 Blackwell Server Edition, DGEMM speedups range from 10x to 23x because its native FP64 is weak.
- If required precision exceeds the default or a user cap, cuBLAS falls back to native FP64; no application code changes are needed.
The FP64 problem
Scientific simulation runs on double precision. Quantum chemistry, materials science and many physics codes spend most of their time multiplying large matrices of 64-bit numbers. AI hardware, meanwhile, has raced toward lower precision, with integer and low-bit Tensor Cores far outpacing native FP64 units. That leaves supercomputing centers in a bind: buy chips built for AI, and the double-precision math science relies on improves slowly.
Nvidia’s answer is emulation. Its cuBLAS library splits each double-precision number into integer pieces, multiplies them on the fast integer Tensor Cores, and reassembles an exact result. CUDA Toolkit 13.4.2 adds two algorithms that make this cheaper: Ozaki-II, which uses the Chinese Remainder Theorem to need far fewer integer multiplications than the earlier Ozaki-I, and a 2M method that cuts a complex matrix multiplication from four real multiplications to two. It builds on the Automatic Dynamic Precision framework Nvidia introduced earlier.
Numbers that matter to scientists
Nvidia reports DGEMM with Ozaki-II runs 5.5x faster than native FP64 on Vera Rubin NVL72 and 3.9x faster on GB200, and twice as fast as Ozaki-I on GB200. Complex ZGEMM with 2M reaches 5x over native on GB200. Peak DGEMM throughput reaches 175 TFLOPS on VR NVL72. Rubin adds a TI16 integer format and a new instruction aimed at exactly these workloads, which shows Nvidia is designing silicon around emulation rather than bolting it on.
The application results are more convincing than kernel benchmarks. CCSD coupled-cluster calculations in quantum chemistry run 2x faster end to end on VR NVL72 and 1.4x on GB200 in full FP64, and 5x and 3.5x in a domain-specific mixed-precision mode. BerkeleyGW, a materials-science code, runs 6.6x faster on VR NVL72 and 3.9x on GB200. On RTX PRO 6000 Blackwell Server Edition, where native FP64 is especially weak, DGEMM speedups run from 10x to 23x. At small matrix sizes on GB200, cuBLAS decides emulation isn’t worth it and uses native FP64.
Trust, and a strategic shift
The obvious worry is accuracy. Nvidia says the framework is designed to give accuracy comparable to or better than standard floating-point GEMM, and if the required mantissa bits exceed the default or a user-set maximum, cuBLAS falls back to native FP64. Users can also set the emulation strategy and bit width manually. That safety valve matters: emulation that silently degrades results would be worse than slow math. Scientists should still validate on their own workloads, which Nvidia explicitly invites.
The strategic signal is clear. Nvidia is telling the HPC community that future GPUs will lean on integer throughput and software to deliver double precision, rather than scaling native FP64 hardware. That works for dense matrix multiplication, which emulation suits. It’s less clear for codes dominated by other FP64 operations. It fits the SC26 story of Rubin HPC and two decades of CUDA in science: Nvidia wants science on AI hardware, and software is the bridge.
Who benefits
- Computational chemists and materials scientists
- Large speedups on DGEMM- and ZGEMM-heavy codes with no code changes.
- Supercomputing centers
- AI-oriented GPUs become more viable for traditional FP64 workloads.
- Workstation users
- RTX PRO 6000 class cards gain 10x to 23x on DGEMM despite weak native FP64.
- Codes beyond matrix multiply
- Workloads dominated by non-GEMM FP64 math see less benefit.
What to watch
- Independent accuracy validation of Ozaki-II across diverse scientific codes.
- Adoption in major HPC applications beyond CCSD and BerkeleyGW.
- Whether rival vendors ship comparable FP64 emulation.
- How Rubin’s TI16 format performs in production supercomputers.
FAQ
Is emulated FP64 as accurate as native?
Nvidia says its framework is designed to give comparable or better accuracy, and it falls back to native FP64 when needed precision exceeds limits.
Do applications need to change code?
No. cuBLAS accelerates DGEMM and ZGEMM automatically, though users can set strategies and bit widths manually.
Which GPUs benefit most?
GPUs with high integer Tensor Core throughput relative to FP64; RTX PRO 6000 Blackwell Server Edition sees 10x to 23x DGEMM gains.
Companies: NVIDIA
More on NVIDIA
AI · Nov 5, 2026
NVIDIA’s four open Cosmos 3 Nano models let robot and driving teams simulate worlds that reactAI · Nov 16, 2026
Rubin HPC is Nvidia’s bet that science agents will be bottlenecked by simulation, not reasoningAI · Nov 9, 2026
Nvidia’s Korean startup bets span interconnect, open models, robot hands and document AIAI · Nov 8, 2026
Nvidia’s Dynamo recipes speed four Korean open models to production, tuned for B200s