What we know

At Supercomputing 2026, AMD highlighted the advantages of its Instinct MI430X accelerator’s native FP64 (double-precision floating point) support. The company criticized Nvidia’s approach, which relies on Ozaki emulation schemes for FP64, stating these schemes cover less than 10% of high-performance computing (HPC) workloads and can triple memory usage. AMD argues that native FP64 support, as provided by the MI430X, is more efficient and better suited for most HPC applications. AMD’s Instinct MI430X accelerator delivers native FP64 support. Nvidia employs Ozaki FP64 emulation, which fits fewer than 10% of HPC workloads and increases memory consumption significantly. Native FP64 support is positioned by AMD as superior for the majority of HPC workloads compared to emulated FP64.

Why it matters

Double-precision floating point performance is essential for many scientific and HPC tasks. Nvidia’s emulation approach can increase memory demands and potentially reduce efficiency. AMD’s native FP64 support on the MI430X could offer improved performance and resource usage for a wider range of HPC applications, which may influence choices in scientific computing hardware.

What is still unknown

No independent evaluations of the MI430X’s FP64 accuracy or performance have been published. Details about supported programming languages, software ecosystem compatibility, and the accelerator’s availability timeline remain undisclosed.