Skip to main content
v2026.11,986 entries · CC-BY 4.0

Editorial · CASRAI · Research data infrastructure

AMD details Instinct MI430X GPU for HPC and AI science

AMD’s 12 November 2026 post details the MI430X FP64 GPU and names Discovery, Alice Recoque and Herder as systems using it.

Published 15 Nov 2026· 4 minute read

On 12 November 2026, AMD published a blog post, “Exascale Was the Beginning: AMD Instinct MI430X GPUs”, describing the MI430X, an accelerator it says was designed from the outset for high-precision (FP64) scientific computing combined with AI. The post coincides with the SC26 supercomputing conference, where AMD researchers presented a paper on FP64 emulation. For university and national-lab computing groups, the practical points are which systems are named, which figures are AMD projections, and when hardware can be expected.

What AMD states about the MI430X

AMD describes the MI430X as built on its CDNA 5 architecture, pairing up to 288 TFLOPS of peak native FP64 with 432 GB of HBM4 memory at 23.3 TB/s and 9.2 PF of peak FP4 performance. It says the node design connects four GPUs to a 6th Gen EPYC “Venice” host in a glue-less Infinity Fabric mesh, with up to 384 GB/s unidirectional per GPU. AMD also says the MI430X keeps the shared-memory programming model of its MI300A, so existing code and expertise carry forward.

These are vendor figures. AMD’s own footnotes say the MI430X results are projections based on engineering estimates and preliminary silicon targets as of November 2026, not measured results, and that they are subject to change. Comparisons with Nvidia hardware, including the claim of 50% more memory capacity than Nvidia Vera Rubin, are AMD’s; for Vera Rubin the footnotes cite Nvidia’s preliminary published specifications, and AMD notes it has not verified third-party claims.

Systems and centres named

AMD says the MI430X has been selected for three flagship scientific systems: Discovery at Oak Ridge National Laboratory, described as the successor to Frontier; Alice Recoque, France’s first exascale system, funded by EuroHPC; and Herder at the High-Performance Computing Center Stuttgart (HLRS) in Germany. The post also says LUMI-AI and LUMI-IQ at CSC Finland will deploy MI430X with 6th Gen EPYC, and describes Lux, the first operational DOE Genesis Mission AI-for-science supercomputer, as live at Oak Ridge today. Installed AMD-based systems it cites include Frontier, El Capitan and LUMI; see the CASRAI guide to the El Capitan supercomputer.

HLRS’s own announcement of Herder gives the planning detail for one of these. HLRS says delivery is scheduled for the second half of 2027, with service expected by the end of 2027, replacing its current system Hunter. It says Herder is built on the HPE Cray Supercomputing GX5000 platform with MI430X GPUs and EPYC “Venice” CPUs, and anticipates peak performance of more than seven times Hunter’s. Funding comes from the German Federal Ministry of Research, Technology and Space through the Gauss Centre for Supercomputing and from the Baden-Württemberg Ministry for Science, Research and Arts. HLRS lists computational engineering, physics, chemistry, climate science and biomedicine among the intended research areas.

One figure differs between the two sources: HLRS’s announcement quotes 19.6 TB/s of memory bandwidth per MI430X, while AMD’s November 2026 post and product page give 23.3 TB/s. Planning documents should cite the current vendor figure and its date rather than copying either number without that context.

The FP64 argument and why it matters to researchers

A central theme of the AMD post is native versus emulated double precision. AMD researchers’ SC26 paper, “FP64 Performance and the Limits of Ozaki Emulation on HPC Systems”, profiled 47 workload configurations across 11 applications and benchmarks on MI350X and MI355X GPUs, and then evaluated Ozaki GEMM emulation across three GPU architectures. AMD reports that, in its sample, fewer than 10% of HPC workloads are dense and GEMM-shaped enough for emulated FP64 to add value, and that workloads often described as memory-bound can become compute-bound if native FP64 throughput is reduced. This is an argument by a vendor with a product to sell, based on its own testing; groups that depend on double precision should test their own codes rather than rely on it. For background on the underlying number format, see the CASRAI guide to IEEE 754.

What it means for university and national-lab HPC

  • Timing. The named systems are planned or in procurement; HLRS’s timeline puts Herder in service at the end of 2027, so the MI430X is a planning input rather than hardware most campus clusters can buy today.
  • Porting and reproducibility. AMD points to HIP porting and its ROCm software stack, and says a long-term-support model is planned to give production deployments two years of stability. Teams planning a move should budget for validating numerical results; see computational reproducibility.
  • Allocations and cost recovery. Most researchers will reach such systems through national centres or allocations rather than owning them; the CASRAI guide to NSF ACCESS allocations and campus recharge covers how those routes differ.
  • Energy. HLRS describes direct liquid cooling and a new facility that reuses Herder’s waste heat, a reminder that facility and power planning now shape procurement; see sustainable HPC.

Sources

Ask CASRAI · full access

Ask about this story

Full access: large monthly usage, up to 10 source passages behind each answer, and MCP and API access for Claude Code, Codex and other CLIs. Cancel any time.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →