Nvidia Rubin HPC: 50,000 CUDA cores for science, shipping 2028

Nvidia used SC26 to do something it has not done in years: talk about simulation instead of tokens. The GPU it showed off will not exist until 2028. The software it shipped alongside it exists today.
Rubin HPC is a core-count play, and Nvidia is not shy about it. Announced at the SC26 supercomputing conference in Chicago, Rubin HPC is a new part in the Vera Rubin platform aimed at simulation and data processing rather than matrix multiplication — more than 50,000 CUDA cores on a single GPU, a sparse data acceleration feature to keep those cores fed, and seventh-generation NVLink for GPU-to-GPU traffic. For scale: the standard Rubin GPU carries 898 Tensor cores and somewhere north of 28,000 CUDA cores, and today's NVLink moves 3.6 terabits per second per card, according to Nvidia's announcement. Wccftech, reading Nvidia's die shot, reckons a Rubin HPC tray pairs two Superchips — each a Vera CPU plus two Rubin HPC dies — for more than 200,000 cores per tray.
"For agents tackling science and engineering, faster simulation means faster discovery," Ian Buck, Nvidia's vice president of hyperscale and high-performance computing, said in the same post. "Rubin HPC accelerates the simulations agents use to test ideas and guide their next steps."
Read the announcement closely and the design philosophy is data movement. Simulation workloads — sparse matrices, stencils, sorting, reductions — are gather-heavy: they reach into a huge dataset, pull a handful of scattered values, compute, and repeat at absurd scale. Adding cores without adding memory bandwidth just builds a wider pipe onto the same clogged drain. Nvidia's answer is sparse-data acceleration plus a faster interconnect, and on paper that is the right answer for the workload.
The number Nvidia did not publish is the interesting one: FP64. Nowhere in the Rubin HPC materials is there a double-precision throughput figure. For a chip sold into supercomputing that is a conspicuous hole — FP64 is the currency of climate models, fusion plasmas and molecular dynamics, the workloads where an answer that is merely approximately right is not an answer. Nvidia has spent a decade tuning its data-center roadmap for low-precision AI math, and Wccftech's read is blunt: the company "has not made any meaningful acceleration in FP64 compute."
AMD went the other direction. Its Instinct MI430X is a sibling of the AI-focused MI455X that strips the design back to native double precision: more than 200 FP64 teraflops, 432 GB of HBM4 at 19.6 terabytes per second, and three supercomputers already announced around it — Herder in Germany, Alice Recoque in France and Discovery in the United States, per AMD's own specification post. German outlet heise notes AMD claims more than six times the FP64 throughput of Rubin, which is a comparison that flatters AMD — Rubin is a pure AI accelerator and is not trying to win that contest — but also one that AMD would not be able to make at all if Nvidia had kept the FP64 flank defended.
That is the uncomfortable part of the Rubin HPC story. The chip arrives in 2028, a year or more after the parts that will define the 2027 exascale procurement round. A lab writing a request for proposals this winter cannot put a 2028 GPU in the rack; it buys MI430X, or it buys the Blackwell-and-Vera generation already shipping. Rubin HPC is a credible answer to a question that will already have been answered by the time it lands. Nvidia has run this play before: Rubin CPX, the inference-optimized Rubin variant it announced with much fanfare, quietly came off the roadmap, per Wccftech.
What does ship now is a software stack with real numbers attached. The CUDA-X libraries Nvidia expanded at SC26 are the part a working scientist can use tomorrow, and the headline figure is astronomical. cuPhoton.xBeam, the module that correlates data from the Event Horizon Telescope — a planet-scale array of radio dishes that produces petabytes per observing run — runs that correlation in eight minutes instead of more than a month, per Nvidia's CUDA-X post. Nvidia says it has already correlated the black hole jets J1924 and 3C279 and can reconstruct the Messier 87 black hole in real time, with the whole pipeline running on distributed GPUs or a single DGX Spark — the desktop box we wrote about when a homelabber strapped 36 of them together for 4.6TB of local AI memory.
The rest of the stack reads like a list of bottlenecks being removed one field at a time. cuPhoton.xLens lets Pacific Fusion inspect plasma simulations while they run instead of after, which is the difference between steering an experiment and reading its autopsy. cuPotential, built on work from the Simons Foundation's Flatiron Institute and D. E. Shaw Research, accelerates existing molecular dynamics engines by up to 3x, with Schrödinger, AMBER, NAMD, GROMACS, LAMMPS and the Open Molecular Software Foundation communities integrating it and Cadence, Fathom Therapeutics, Topos Bio, Genesis Molecular AI and Terray Therapeutics adopting it. Daviator brings VASP calculations into Nvidia's ALCHEMI stack at up to 22x the speed and 75% lower cost. cuSpa, now open source, computes cell polygons 900x faster than CPUs for spatial models at Memorial Sloan Kettering. The GPU-accelerated HMMER library runs protein annotation up to 84x faster at Basecamp Research. Post-training gets its own kit: Nemotron for Science, built on the open-source Nemotron 3.5 Super VL — 124 billion parameters, roughly a tenth active at a time, partly Mamba-2 — plus pre-packaged post-training recipes and nearly 50,000 chemistry questions, per SiliconANGLE.
There is a quantum thread too. Nvidia added building blocks and a qubit error-mitigation algorithm to CUDA-Q, and is working with IBM and the Department of Energy's Fermilab on a modular quantum codesign platform that pairs DGX B200 systems with IBM quantum hardware.
The honest caveat is that every one of those multipliers compares Nvidia to a CPU, not to AMD or to last year's Nvidia. An 8-minute correlation is a genuine result if it holds up in someone else's hands; it is a press slide if it does not. The same goes for the "agents for science" framing that runs through all of Nvidia's SC26 material: an agent that proposes a molecule is not evidence, a paper that survives review is. What the software announcement really does is lock in the toolchain — a graduate student who learns cuPhoton and cuSpa this year has little reason to port anything to ROCm next year, and that habit is worth more to Nvidia than any single benchmark. The chip is the marketing; the libraries are the moat.
What to watch: whether Nvidia publishes an FP64 figure for Rubin HPC before SC27, whether Herder, Alice Recoque and Discovery stay on schedule with MI430X, and whether any of these speedups show up in a peer-reviewed methods section this winter. Nvidia reports quarterly results today, and none of the SC26 material changes the data-center revenue story — but the HPC segment is the one place where a competitor still has a genuine opening.
Would you rather have the faster chip in 2028 or the working software today? Write to us via our contact page.







