Nvidia announced Rubin HPC at SC26 on 16 November 2026, adding a GPU designed for scientific simulation and data processing to its Vera Rubin platform. The company says the chip will have more than 50,000 CUDA cores and is intended for workloads in which simulations repeatedly draw values from large datasets. Its pitch is aimed partly at scientists and engineers using AI agents to propose materials, molecules and designs that then need testing.
Key points
- Rubin HPC is a simulation-focused GPU in Nvidia’s Vera Rubin platform.
- Nvidia says it will combine more than 50,000 CUDA cores with sparse data acceleration and seventh-generation NVLink.
- The company expects the GPU to be available in 2028.
More than 50,000 cores need data
A CUDA core performs calculations on the GPU, but the number of cores alone cannot determine how quickly a simulation finishes. Nvidia describes the scientific workloads targeted by Rubin HPC as a repeating sequence: retrieve a small selection of values scattered through a large dataset, calculate with them, then do it again at scale. The company says Rubin HPC is designed to increase data movement alongside computing capacity so its cores spend less time waiting for inputs.
Finding those values is closer to pulling particular pages from many folders than reading one book from beginning to end. Nvidia calls the chip’s provision for this sparse data acceleration. It targets recurring operations such as sparse matrix calculations, which work with collections containing many empty entries, as well as stencil computations, sorting and reductions. The company says these patterns occur across climate modelling, engineering simulation, genomic analysis and financial risk workloads.
Testing a design before making a physical prototype could involve examining more candidate designs if the simulations used to assess them returned results sooner.
Nvidia also specifies seventh-generation NVLink, the connection that lets GPUs exchange data. Moving values efficiently within one GPU and passing data between GPUs are separate parts of the design, both relevant to calculations spread across multiple chips. Ian Buck, Nvidia’s vice president of hyperscale and high-performance computing, said the intended gain is faster feedback for agents testing ideas through simulation.
Rubin HPC joins Vera Rubin systems
Rubin HPC is a variant within Vera Rubin rather than the name of an entire new platform. Existing configurations include NVL72, with 72 Rubin GPUs and 36 Vera CPUs, and HGX Rubin NVL8, which connects eight Rubin GPUs, The Defiant reported. Those configurations use sixth-generation NVLink, compared with the seventh generation Nvidia has announced for Rubin HPC.
The platform already has a configuration aimed at scientific work: Vera Rubin NVL4 combines four Rubin GPUs and two Vera CPUs, according to The Defiant. Rubin HPC’s stated emphasis is narrower. Nvidia is presenting a chip built around the repeated movement and calculation patterns of simulation and analytics, with the prospect of agents requesting more such runs as they explore possible materials, molecules and designs.
For developers building those workflows, Nvidia points to CUDA-X, its libraries of optimised algorithms for simulation and analytics. The company says agents could use the libraries with Rubin HPC to run simulations that inform their next steps. That proposal covers both the hardware doing the calculations and software implementations of the operations those applications repeatedly call.
Nvidia’s 2028 availability expectation
Nvidia’s announcement gives a core count and describes how the chip is intended to move data, but The Defiant reported that it does not provide an application benchmark, memory capacity or double-precision arithmetic performance. An application benchmark would measure a complete workload rather than an individual design feature. The publication also reported that Nvidia had not identified the board or complete system in which Rubin HPC would ship.
Nvidia expects Rubin HPC to be available in 2028.