B200 complex matrix multiplication up to 295 TFLOPS… NVIDIA supports Ozaki-II
Original headlineB200 복소 행렬곱 최대 295TFLOPS…엔비디아, Ozaki-II 지원
Not investment advice. Muhabbit scores how much a story may matter, not which way to trade.
NVIDIA added the Ozaki-II FP64 matrix-multiplication emulation path to cuBLAS starting with CUDA Toolkit 13.4, supporting Ampere and later GPUs including the B200, where complex matrix-multiply emulation peaks at 295TFLOPS. NVIDIA said the new path is used when it outperforms the earlier Ozaki-I method, which first arrived as an optional feature in CUDA Toolkit 13.0 Update 2. The 295TFLOPS figure is a best-case result for a specific compute path, not a guarantee that full scientific simulations will run at that rate. The change matters because FP64 precision underpins scientific and engineering workloads, and emulation offers a way to recover double-precision throughput on hardware that prioritizes lower-precision AI math. What to watch next is independent benchmarking of Ozaki-II against native FP64 and Ozaki-I across real simulation workloads, and whether developers adopt it by default.
Muhabbit's AI wrote this summary in its own words from TokenPost's report and translated it from Korean. It can get details wrong, so check the source.
Named in this story
NVIDIA (NVDA), B200 GPU, CUDA Toolkit, Ozaki-II, cuBLAS
A linked name opens that asset's sentiment page.