{
  "id": 20530793,
  "title": "GPU giants throw HPC users some tasty crumbs",
  "url": "https://urgent.news/2026/11/17/gpu-giants-throw-hpc-users-some-tasty-crumbs-20530793",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-11-17T22:37:46.000Z",
  "source": {
    "name": "The Register",
    "slug": "the-register",
    "url": "https://www.theregister.com/hpc/2026/11/17/gpu-giants-throw-hpc-users-some-tasty-crumbs/5310611"
  },
  "original_language": "en",
  "account": "At Supercomputing 2026, Nvidia unveiled the Rubin HPC variant of its Rubin GPUs, tailored for high performance computing (HPC). The chip offers over 50,000 CUDA cores, nearly double that of the AI-focused version introduced earlier this year. According to Nvidia's senior director of supercomputing, Dan Ernst, Rubin HPC features a rebalanced architecture optimized for simulation, engineering, and high-performance analytics workloads. Although Nvidia has not disclosed peak performance memory capacity or bandwidth, the chip is likely to include features demonstrating its commitment to HPC beyond its AI focus. Concerns have emerged that Nvidia could disregard HPC, as seen in the shift from Blackwell Ultra's double-precision matrix acceleration towards higher throughput at lower precisions. Rubin's 33 teraFLOPS performance rate is slower than its predecessors and even older Hopper accelerators. However, Nvidia claims that the higher memory bandwidth-to-compute ratio for sparse vector-heavy workloads measured by the High Performance Conjugate Gradients (HPCG) benchmark should yield meaningful performance gains. For dense matrix-heavy workloads, Nvidia utilizes FP64 emulation, employing an algorithm derived from the Ozaki Scheme to decompose FP64 matrix operations into INT8 operations, achieving higher-than-native performance. Rubin can reportedly execute up to 200 emulated FP64 matrix teraFLOPS. AMD is also developing HPC-chip-tuned GPUs, with the MI400-series Instinct MI430X boasting 432 GB of HBM4, 23 TB/s of memory bandwidth, and 288 teraFLOPS of peak double precision performance. AMD's chips leverage a modular chiplet architecture to build distinct versions for AI and HPC tasks. Performance improvements are anticipated for top HPC systems such as EuroHPC's Alice Recoque and Oak Ridge National Laboratories' upcoming Discovery.",
  "summary": "Nvidia unveils an HPC-optimized Rubin GPU to challenge AMD's 288 teraFLOPS MI430X",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}