NVIDIA Says FP64 Emulation Lifts Rubin Matrix Math to 175 TFLOPS; Vector Code Sees No Gain
Reported by Tech Times and Tom's Hardware2 sources ↓
NVIDIA says its cuBLAS FP64 emulation reaches about 175 TFLOPS on Vera Rubin NVL72, up to 5.5 times native FP64 throughput in its tests. AMD argues emulation does not replace native FP64 for many HPC workloads.
Key points
- NVIDIA reports Ozaki-II delivers about 175 TFLOPS on Rubin NVL72, versus about 33 TFLOPS for native FP64
- The reported gain applies to matrix multiplication, not vector-heavy workloads
- NVIDIA says its library can fall back to native FP64 when emulation is not safe
- AMD says fewer than 10% of HPC workloads perform dense GEMM operations
- AMD says MI430X offers native FP64; its cited systems include Discovery, Alice Recoque, LUMI-AI and Herder
Sources
Details
NVIDIA's reported results are company benchmarks, and no independent grading of the new Rubin emulation path was cited. The Tech Times report says NVIDIA reports up to 3.9 times native throughput on GB200 Blackwell, and gains of roughly 10 to 23 times on RTX PRO 6000 Blackwell Server Edition, depending on matrix shape. Those comparisons use different hardware and baselines.
The method, called Ozaki-II, breaks FP64 values into residues using the Chinese Remainder Theorem, performs integer multiplications, then reconstructs the result. NVIDIA says cuBLAS's Automatic Dynamic Precision estimates the precision needed, adds less than 10% overhead, and uses native FP64 when emulation is not safe. The report notes concerns that some input distributions can reduce accuracy and that the new Rubin path has not received independent grading.
The gain is limited to general matrix multiplication. Tech Times reports that NVIDIA says cuBLAS selects native FP64 when problems are too small to benefit, while vector-heavy codes see no emulation speedup. AMD, as reported by Tom's Hardware, says fewer than 10% of HPC workloads perform dense GEMM and that emulation can require up to three times the memory capacity of native FP64; Tech Times separately reports AMD researcher Nicholas Malaya describing about twice the memory use.
Written from coverage as of
Images
- Tom's Hardware
- Tom's Hardware
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh