AMD attacks Nvidia's FP64 emulation at SC26, touts native double-precision on Instinct MI430X — says Ozaki schemes fit l

Semiconductors

AMD Says Nvidia's FP64 Emulation Fits Under 10% of HPC Workloads

AMD says Nvidia's Ozaki FP64 emulation fits under 10% of HPC workloads and can triple memory use, pitching the 432GB HBM4 Instinct MI430X as the native double-precision alternative at SC26.

By
Rebecca Stone
Filed
Channel
Semiconductors
Read
4 min read

AMD claims that less than 10% of HPC workloads perform the dense GEMM operations that Nvidia's Ozaki-scheme FP64 emulation depends on — and that emulation can triple an application's memory footprint compared with native double precision. The company laid out its case this week as Supercomputing 2026 opened in Chicago, positioning its Instinct MI430X accelerator as the answer for national laboratories and supercomputing centers that still need real FP64 hardware.

Nvidia, over the past few years, has reallocated die area on its data center GPUs toward the lower-precision matrix math that drives most modern AI inference. The logic is straightforward. The addressable market for AI compute has exploded, while demand for double-precision floating-point math in large supercomputing installations has stayed essentially flat.

To fill the gap, Nvidia turned to emulation techniques, including the Ozaki schemes. AMD says that's not a substitute.

Why does AMD think Ozaki emulation falls short?

The core objection is workload shape. Ozaki emulation, AMD notes, assumes the workload resembles a dense general matrix multiply. Based on its own application profiling, the company found that fewer than one in ten HPC workloads actually perform dense GEMM operations. Dense linear algebra is the one common case where AMD concedes Ozaki emulation is a natural fit.

Everything else is a problem. For sparse iterative linear algebra, stencils and hydrodynamics, high-order finite element analysis, Monte Carlo and particle transport, and classical molecular dynamics, AMD says developers still need strong FP64 vector performance for the best results.

Even GEMM-friendly workloads carry caveats. Matrix sizes must be large to hit peak performance, AMD says, and that isn't a given for every operation. Operands also have to be sized correctly to the hardware; incorrectly scaled operands drag performance down. The FP64 vector units in the MI430X, by contrast, don't care about operand sizing.

Then there's memory. AMD says Ozaki emulation can balloon an application's memory footprint to as much as three times what native FP64 requires.

What does that mean for system design?

HBM capacity is welded directly to GPUs in modern supercomputing installations. Squeeze a workload onto hardware that only emulates FP64, and the cascading effects start: more GPUs per system, more compute nodes, more floor area, more power, more cooling.

AMD's counter is the MI430X node, which carries 432GB of HBM4 memory. That capacity, the company says, lets customers provision fewer GPUs and fewer nodes, cutting costs on networking, power, and floor space.

Can an MI430X box do AI too?

Yes, but within limits. AMD pitches an MI430X-powered supercomputer as a platform for AI inference in research workflows, alongside the double-precision compute. The numbers make the tradeoff clear: each MI430X GPU delivers 9.2 PFLOPS of theoretical FP4 compute for inference, against 40.3 PFLOPS of FP4 from each AI-focused MI455X accelerator in the same CDNA 5 architectural family.

Scale-up domains differ too. Four MI430Xs link over AMD's Infinity Fabric alongside a sixth-generation EPYC host CPU — a domain that pales next to the 72 MI455X accelerators in a Helios rack. AMD says the MI430X interconnect is optimized for latency-sensitive operations rather than a massive scale-up domain.

Still, the node holds its own for scientific AI. Four MI430Xs working together offer 1.72TB of aggregate HBM4 capacity, 93.2 TB/s of memory bandwidth, and 12.3 Tbps of scale-up bandwidth within the node — likely more than enough, AMD argues, for the AI applications that need to run alongside double-precision workloads. Institutional supercomputing installations, after all, aren't usually trying to double as cloud inference providers.

Where has the MI430X already landed?

The accelerator already has several supercomputing design wins. The upcoming Discovery system at Oak Ridge National Laboratory, the Alice Recoque installation in France, the LUMI-AI system at Finland's CSC, and the Herder system in Germany will all use the MI430X's native FP64 support.

That track record is the point. Nvidia has retreated from native double precision in its data center products, and AMD has kept building for it. The company's semi-custom silicon design experience and chiplet technology give it the flexibility to tailor products like the MI430X to specialized markets rather than chase a one-size-fits-all design. Those institutional machines will be dwarfed by the AI inference installations that AMD and Nvidia are shipping elsewhere — but for the labs that need FP64, and need it without a 3x memory penalty, AMD is betting the MI430X remains the only serious option on the floor.

Source: Tom's Hardware

Share this article:

More from Rebecca Stone

Rebecca Stone

Show full bio

Correspondent covering technology and product news in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,092 articles

Related articles

« Previous articleNext article »