Skip to content
audiotale.net
Sign in
NewsNews

NVIDIA Rubin HPC GPU: more than 50,000 CUDA cores, due 2028, plus a four-node DGX Spark kit

NVIDIA unveiled Rubin HPC at SC26 with a record CUDA core count but no memory or FP64 figures, and launched a 512GB four-node DGX Spark developer kit.

audiotale.net team

· 6 min read

Written by AI from the sources linked below. How we use AI

On this page
  1. Key points
  2. What happened
  3. Background
  4. The numbers
  5. What else NVIDIA announced at SC26
  6. What it means
  7. What we don't know yet
  8. What to watch next
  9. Sources

NVIDIA used the opening of the SC26 supercomputing conference to announce Rubin HPC, a new member of its Vera Rubin GPU family that the company says has "more than 50,000" CUDA cores on a single GPU, the most it has put in any chip. It will not ship until 2028, and the company left out the numbers that high-performance computing buyers care about most: memory capacity and bandwidth, and how much double-precision (FP64) performance the chip really has. Alongside it, NVIDIA announced a four-node DGX Spark developer kit for researchers and a Slurm update that lets supercomputers schedule quantum processors.

Key points

  • NVIDIA announced Rubin HPC on 16 November 2026 at SC26; it says the GPU has more than 50,000 CUDA cores and uses seventh-generation NVLink.
  • NVIDIA says Rubin HPC is expected to be available in 2028; it has not published memory, bandwidth, power or FP64 figures for it.
  • Tom's Hardware notes today's Rubin GPU has 224 streaming multiprocessors and 28,672 CUDA cores, so Rubin HPC would have at least about 1.7 times as many cores.
  • The DGX Spark HPC Developer Kit links four DGX Spark systems in a ring for 512GB of combined memory and 20TB of NVMe storage, and is on sale now through partners.
  • Slurm 26.11, due later this month, adds support for quantum processors; NVIDIA says nearly 30 supercomputing centres are deploying more than 50 of them.

What happened

NVIDIA's announcement describes Rubin HPC as a new GPU in the Vera Rubin platform designed to accelerate simulation and data processing for science, engineering and analytics. The pitch is built around agents: NVIDIA argues that as AI agents explore more candidate materials, molecules and designs, they lean on simulation to test those ideas, so faster simulation shortens the loop from hypothesis to result. Ian Buck, NVIDIA's vice president of hyperscale and high-performance computing, put it in the release as "faster simulation means faster discovery".

The hardware claims are short. Rubin HPC has a "record number" of CUDA cores, more than 50,000 on one GPU. It adds a sparse data acceleration capability meant to keep those cores supplied with data, and it uses NVLink 7 for GPU-to-GPU connections. NVIDIA calls the design "balanced": it says it scales data movement together with compute instead of raising compute alone. That, the company argues, suits workloads in climate models, engineering simulation, genomic analysis and financial risk that repeatedly reach into a very large dataset, gather a few scattered values, compute a result and repeat. NVIDIA lists sparse matrix operations, stencil computations, sorting and reductions as the recurring patterns it targets, and says its CUDA-X libraries will provide tuned implementations.

Tom's Hardware, which covered the announcement, points out how much was left unsaid. NVIDIA did not disclose Rubin HPC's HBM capacity or bandwidth, did not give new NVLink 7 bandwidth figures, and did not say whether the chip shifts more of its compute toward native FP64, as AMD and Intel are promoting for their own GPUs at the same conference. Tom's Hardware also infers that reaching more than 50,000 cores would require more dies on a package, perhaps similar to the four-die Rubin Ultra design, but notes NVIDIA did not describe the layout. That is the outlet's inference, not an NVIDIA statement.

Background

Rubin HPC follows an announcement from ISC High Performance in June 2026, where NVIDIA said the Vera Rubin platform could deliver more than 7 exaflops of AI performance for science and 5 petaflops of native FP64 in a full system with up to 144 GPUs, which it said puts a rack on par with systems on the TOP500 list. Those are NVIDIA's own system-level claims, not measured TOP500 results. At the time it named three supercomputers built on Vera Rubin: Blue Lion at the Leibniz Supercomputing Centre (scheduled for 2027, with about 30 times the computing power of the centre's current system), Doudna at NERSC and the Lawrence Berkeley National Laboratory, and new systems at Los Alamos National Laboratory. Vera Rubin NVL4 systems from partners including Dell, HPE and Supermicro were expected in the fourth quarter of this year.

Seen against that, Rubin HPC is a second, science-focused GPU rather than the first Rubin for science. It is a 2028 part, so it matters for purchasing roadmaps now but not for any machine you can order today.

The numbers

The one hard comparison available is core count. Tom's Hardware gives today's Rubin GPU 28,672 CUDA cores across 224 streaming multiprocessors, and NVIDIA's own figure for Rubin HPC is "more than 50,000", so the chart shows a floor rather than a final number.

CUDA cores per GPU (Rubin HPC is a stated minimum)CUDA cores
CUDA cores per GPU (Rubin HPC is a stated minimum) (CUDA cores)
CUDA cores
Rubin HPC (more than)50000
Rubin (current)28672

Source: NVIDIA blog (Rubin HPC); Tom's Hardware (current Rubin)

Core count alone says little about delivered performance. HPC workloads such as sparse solvers are often limited by memory bandwidth, not arithmetic, which is why NVIDIA stresses the sparse data acceleration and interconnect. Until NVIDIA publishes bandwidth, memory size and FP64 rates, the 50,000 figure is a design indicator, not a benchmark.

What else NVIDIA announced at SC26

DGX Spark HPC Developer Kit. The kit is four DGX Spark systems, each built on the GB10 Grace Blackwell Superchip, connected directly to two neighbours in a ring over ConnectX-7 networking, so no switch is needed. At four nodes it offers 512GB of combined memory and 20TB of NVMe storage. NVIDIA says that with NVFP4 precision researchers can run models of up to 700 billion parameters across the four systems, naming DeepSeek V4.1 as an example. An NVIDIA HPC SDK installation playbook is "coming soon". The kit is on sale now, with parts from OEM partners including Acer, ASUS, Dell, GIGABYTE, HP, Lenovo and MSI. NVIDIA cites the Rubin Lab at IGI at UC Berkeley, which reports results arriving 15 times sooner than waiting in an 11-hour cluster queue. That is a customer claim published by NVIDIA.

Slurm and quantum. NVIDIA says Slurm, the scheduler many supercomputing centres use, is getting support for quantum processors in release 26.11, due later this month. The aim is to let hardware makers write one description of their device that works across centres. NVIDIA says nearly 30 centres are deploying more than 50 quantum processors, with adopters including IBM, IQM, Pasqal and Quantinuum.

Co-packaged optics in production. ServeTheHome reports that NVIDIA and cloud provider Lambda disclosed what it calls the first production cluster using co-packaged optics, in a deployment of more than 10,000 GPUs. The approach moves optical transceivers onto the switch package, which can cut power use compared with pluggable optics. This is the outlet's account of the disclosure, and we did not find Lambda's own figures.

What it means

For research computing teams, Rubin HPC is a signal about direction. NVIDIA is putting its weight behind agent-driven science and is positioning general-purpose GPUs as the instrument for it, while AMD and Intel pitch native FP64 strength. The competitive question for buyers who run classic double-precision codes is whether Rubin HPC keeps strong FP64 or leans on lower-precision and emulation, and NVIDIA has not answered that yet.

For developers and students, the DGX Spark kit is the nearer-term news. It gives a small team a four-node cluster to practise MPI-style multi-node programming and run large quantised models locally, without cluster queues, though 700 billion parameters at NVFP4 is NVIDIA's stated limit, not a benchmark we have seen run.

For the industry, the 2028 date suggests NVIDIA is planning HPC parts on a longer cadence than its AI racks, and the Slurm work shows quantum processors being treated as one more resource in a scheduler rather than a separate world.

What we don't know yet

  • The memory type, capacity and bandwidth of Rubin HPC, its FP64 and FP32 rates, power draw and package layout.
  • The price of the DGX Spark HPC kit; NVIDIA's post points to partners and its marketplace.
  • How the 4-node ring performs on communication-heavy codes compared with a switched cluster.
  • Whether the 2028 date holds. NVIDIA's release carries the usual forward-looking caveats.

What to watch next

Watch for NVIDIA's first detailed Rubin HPC specifications and for the FP64 numbers AMD and Intel are promoting at SC26. Also watch the NVIDIA HPC SDK playbook for the DGX Spark kit and the Slurm 26.11 release later this month.

Sources

Written by

audiotale.net team

The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.