Wednesday, November 18, 2026
AI 인프라 · 뉴스 & 분석
홈 › 반도체·하드웨어 › 리포트
반도체·하드웨어 · 리포트

At SC26, AMD touted native FP64 double-precision support on its Instinct MI430X accelerator while attacking Nvidia's FP64 emulation approach.

For HPC and scientific workloads, native FP64 hardware could strengthen AMD's pitch in supercomputing procurement against Nvidia's software-emulated alternative.
업계 전문지Slicast · 2026년 11월 17일 14:30 UTC · 글로벌 · 출처: Tom's Hardware
중요도 50

As Supercomputing 2026 (SC26) opens in Chicago this week, AMD is indirectly needling Nvidia over its lack of native FP64 data type support in its data center GPUs. FP64, or double-precision floating-point math, is widely used in the high-performance scientific computing and simulation fields that are the focus of SC26. For those customers, AMD is promoting its latest Instinct MI430X accelerator, aimed at customers that still need native FP64, such as national laboratories and supercomputing centers.

Over the past few years, Nvidia has reallocated die area on its data center GPUs to favor the lower-precision matrix math that boosts the operations needed for much of modern AI inference. That reallocation makes business sense, as the total addressable market for AI compute has exploded while demand for double-precision math in large supercomputing installations has remained flat by comparison.

To make up for the lack of native FP64 support, Nvidia has turned to techniques that emulate FP64 operations using methods such as the Ozaki schemes. AMD claims this emulation is nowhere close to a replacement for native double-precision processing, for several reasons, and that is why it continues to develop specialized products like the MI430X.

For one, AMD notes that Ozaki emulation assumes the workload resembles a dense general matrix multiply (GEMM). In its own application profiling, the company says "less than 10%" of HPC workloads perform dense GEMM operations, so emulation is a good replacement for native compute only for that small share. Dense linear algebra is the one common workload for which AMD says Ozaki emulation is a natural fit.

For workloads such as 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 to get the best results.

Even for workloads that map well onto GEMM operations, AMD says matrix sizes must be large to achieve the best performance, which may not hold for every operation. The operands of those matrices must also be sized well to the hardware, since incorrectly scaled operands can reduce performance. The FP64 vector math capabilities of the MI430X are insensitive to these operand sizing issues.

Beyond those hurdles, AMD notes that Ozaki emulation can cause an application's memory footprint to balloon, requiring up to three times more capacity than native FP64 would. Because HBM capacity is directly tied to GPUs in modern supercomputing installations, fitting a workload onto a machine that can only emulate FP64 could have cascading effects on the number of GPUs and compute nodes required. That, in turn, creates knock-on challenges for the floor space, power, and cooling needed to support the system.

AMD says the 432GB of HBM4 memory capacity per MI430X node allows customers to provision fewer GPUs and compute nodes in a system, saving on networking, power, and floor space.

In addition to its FP64 HPC performance, AMD positions an MI430X-powered supercomputer as an AI inference platform for developing workloads that use those models as part of the research process. There are key differences, however, between those systems and the Helios rack-scale accelerator powered by MI455X GPUs from the same CDNA 5 architectural family.

Each MI430X GPU offers 9.2 PFLOPS of theoretical FP4 compute for inference, which should be understood in the context of the 40.3 PFLOPS of FP4 compute from each AI-focused MI455X accelerator.

Beyond raw performance, the native scale-up domain of four MI430Xs over AMD's Infinity Fabric interconnect, paired with a sixth-generation EPYC host CPU, pales in comparison to the 72-accelerator MI455X scale-up domain in a Helios rack. AMD says this interconnect is optimized for latency-sensitive operations rather than for a massive scale-up domain.

Of course, every customer will have different requirements, and institutional supercomputing installations usually aren't trying to double as cloud inference providers. For scientists and engineers whose AI models of choice fit within the 1.72 TB of aggregate HBM4 capacity and 93.2 TB/s of memory bandwidth that four MI430X GPUs offer working together, the 12.3 Tbps of scale-up bandwidth available within such a node should be more than sufficient for the AI applications that run alongside double-precision compute workloads.

In any event, the MI430X already has several supercomputing design wins. The upcoming Discovery supercomputer 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 take advantage of the native FP64 support the MI430X offers.

AMD's continued support for native FP64, even as Nvidia retreats from that capability in its products, will doubtless be a point of pride for its architects and engineers, even though these institutional supercomputers will be dwarfed by AI inference installations powered by both AMD and Nvidia hardware around the world. The company's semi-custom silicon design expertise and chiplet experience give it the flexibility to tailor products like the MI430X to these specialized markets rather than pursuing a one-size-fits-all approach.

원문 보기
At SC26, AMD touted native FP64… · Slicast