NVIDIA has announced Rubin HPC, a new GPU in its Vera Rubin platform, designed to accelerate simulation and data processing for science, engineering and analytics workflows. The company made the announcement at SC26 on 16 November 2026. It expects the GPU to be available in 2028 and gave no price.

What NVIDIA says about the design

NVIDIA says Rubin HPC has a record number of CUDA cores, with more than 50,000 on a single GPU. That figure is the company's own claim, and no independent measurement exists yet. The company also says the GPU includes a sparse data acceleration capability intended to keep those cores supplied with data, but it gave no figures for it.

According to NVIDIA, the design is balanced, scaling data movement and compute together rather than compute alone. It says the target is recurring patterns in scientific code, such as sparse matrix operations, stencil computations, sorting and reductions.

The confirmed interconnect detail is that Rubin HPC uses seventh-generation NVIDIA NVLink. The announcement material as summarised to us does not give the NVLink bandwidth.

Agents and simulation

NVIDIA positions the GPU around AI agents that run simulations to test candidate materials, molecules and designs. It says its CUDA-X libraries provide optimised implementations of algorithms for agents to use alongside the hardware.

For agents tackling science and engineering, faster simulation means faster discovery

Ian Buck, NVIDIA vice president of hyperscale and high-performance computing

On the same day, NVIDIA introduced Nemotron for Science, which combines the Nemotron 3.5 Super VL open model, the LitmusBench v0.2 chemistry dataset of nearly 50,000 questions, and post-training recipes. The model and dataset are available now on Hugging Face. That is a separate software announcement from the Rubin HPC hardware.

Why it matters

Our reading: with availability expected in 2028, Rubin HPC is a long-lead roadmap announcement rather than hardware that can be bought or deployed soon. It tells HPC planners where NVIDIA intends to go, but without precision throughput, memory or power figures there is nothing yet to plan procurement around.

What we do not know yet

  • FP64 and other precision throughput figures
  • Memory capacity, type and bandwidth
  • Power draw (TDP) and process node
  • Measured performance compared with current NVIDIA HPC GPUs; no comparative figures were given
  • Price, and which system makers or supercomputing sites have committed to it
  • How Rubin HPC relates to the other Vera Rubin GPUs, and the NVLink bandwidth
  • Independent confirmation: the specifications come only from NVIDIA