
AMD: Nvidia-Style FP64 Emulation Helps Under 10% of HPC Work
AMD's SC26 paper says under 10% of HPC work suits Ozaki FP64 emulation, as it ships the 288-TFLOPS MI430X to Oak Ridge, France and Germany.
- By
- Tom Whitfield
- Filed
- Channel
- AI & Compute
- Read
- 3 min read
AMD told the Supercomputing 2026 conference in Chicago this week that less than 10% of high-performance computing work is dense and matrix-shaped enough to benefit from emulated FP64. The figure comes from an AMD paper that profiled 47 workload configurations across 11 applications, and it lands directly on a technique Nvidia has adopted as it hands more GPU silicon to low-precision AI math.
The finding matters because double-precision floating point, roughly 16 significant decimal digits, remains the standard for climate, physics, engineering and nuclear stockpile simulation, where small rounding errors compound over time. In plain terms, AMD is arguing that most scientific codes cannot borrow the shortcuts built for AI.
What does the emulation actually do?
The schemes in question, known as Ozaki schemes, rebuild an FP64 result from many low-precision tensor-core operations. The approach works well for large dense matrix multiplications, the GEMM kernels at the heart of some simulation codes. Nvidia has leaned on it as FP64 hardware takes a smaller share of its dies.
AMD's paper, titled "FP64 Performance and the Limits of Ozaki Emulation on HPC Systems," found a different picture. After profiling the 47 configurations on its Instinct MI350X and MI355X GPUs and testing emulation on three current GPU architectures, researchers concluded that only one of the 11 applications, a code with a dominant dense matrix kernel, even qualified as a partial fit.
The rest of HPC, AMD said, consists of vector operations, reductions, stencils and sparse linear algebra that cannot be reshaped into one large dense matrix multiply. The company listed the workloads that still need strong native FP64 vector performance:
- Sparse iterative linear algebra
- Stencils and hydrodynamics
- High-order finite element analysis
- Monte Carlo and particle transport
- Classical molecular dynamics
Even where emulation fits, the costs are real. AMD said Ozaki schemes only pay off for large matrices, remain sensitive to how input values are scaled, and can push memory use to roughly three times that of a native FP64 matrix multiply. Memory per GPU is fixed, so a larger footprint can mean more GPUs and more nodes, with knock-on costs in power, cooling and floor space, as Tom's Hardware noted.
Where the MI430X fits
The commercial pitch behind the paper is AMD's Instinct MI430X, built on the CDNA 5 architecture. The company quoted up to 288 TFLOPS of peak native FP64, 432 GB of HBM4 at 23.3 TB/s, and 9.2 petaflops of peak FP4 for AI. Four GPUs link to a sixth-generation EPYC host in each node.
One caveat deserves emphasis: AMD said the MI430X figures are projections from analytical models and preliminary silicon targets, not measured results.
The GPU already has customers. AMD confirmed the MI430X will power Discovery at Oak Ridge National Laboratory, the successor to Frontier; Alice Recoque, France's first exascale system; Herder at HLRS in Germany; and LUMI-AI at CSC in Finland.
How does Nvidia answer?
Nvidia announced its own science-focused GPU, Rubin HPC, at SC26 on Monday. The part carries more than 50,000 CUDA cores and is expected to be available in 2028. Nvidia has not published Rubin HPC's FP64 throughput, StorageReview noted.
The argument therefore splits along a clear line. AMD is selling native double-precision throughput it can name today, against a rival whose science GPU arrives in 2028 with an unspecified FP64 rate and an emulation strategy AMD's own data says serves fewer than one workload in ten.
With Discovery, Alice Recoque and Herder all slated for MI430X hardware, the burden now shifts to Nvidia to show measured FP64 figures, and to laboratories to test whether emulation can stretch their budgets or simply move the cost into memory, power and floor space.
Source: mlq.ai
More from Tom Whitfield
Show full bio
Staff writer covering companies, deals and strategy in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,026 articles
Related articles
amd-mi430x-targets-288-tflops-of-fp64-as-ai-reshapes-hpc
AMD MI430X targets 288 TFLOPs of FP64 as AI reshapes HPC
tensordyne-tapes-out-138-billion-transistor-napier-inference-chip-ee06bfec
Tensordyne Tapes Out 138-Billion-Transistor Napier Inference Chip
amd-buys-taalas-to-hard-wire-ai-inference-into-silicon-5aa48930
AMD Buys Taalas to Hard-Wire AI Inference Into Silicon
microsoft-commits-billions-to-amd-helios-racks-for-azure-inference-1f4c04ad
Microsoft Commits Billions To AMD Helios Racks For Azure Inference


