Nvidia's Rubin HPC GPU Targets Scientific Workflows
Nvidia announced the Rubin HPC GPU at Supercomputing 2026, featuring over 50,000 CUDA cores and NVLink 7 for agentic scientific and engineering workloads, with availability expected in 2028.

Nvidia unveiled the Rubin HPC GPU at the Supercomputing 2026 conference. The company said the processor features "more than 50,000" CUDA cores on a single GPU and uses the new seventh-generation NVLink interconnect technology for communication. Designed for scientists and engineers, the chip targets agentic high-performance computing workloads with a planned release in 2028.
The announcement confirms the GPU as a new member of the broader NVIDIA Vera Rubin platform. Nvidia did not disclose key specifications, including memory capacity, bandwidth, power consumption, or pricing. The company also did not reveal whether the Rubin HPC shifts compute resources toward native FP64 performance.
Technical specifications and design focus
The Rubin HPC GPU's base configuration features 224 Streaming Multiprocessors (SMs) and 28,672 CUDA cores. This balanced design aims to improve data movement and reduce idle compute resources, a critical factor for scientific simulation and data processing. The chip is engineered to accelerate specific workloads like sparse matrix calculations, stencil computations, sorting, and reductions.
It includes dedicated sparse data acceleration to keep its vast array of CUDA cores consistently supplied with data. The GPU will work with Nvidia's suite of CUDA-X libraries. Its architecture is implied to use a four-die layout, similar to the Rubin Ultra GPU. For connectivity, it utilizes NVLink 7, which retains the 3.6 TB/s per-GPU bidirectional bandwidth introduced with the previous NVLink 6 generation.
Platform and workflow integration
Rubin HPC is integrated into the NVIDIA Vera Rubin platform. Its primary design purpose is to serve agentic workflows in scientific, engineering, and analytical domains. Nvidia expects that better agent performance will significantly speed up simulations and the pace of discovery.
The company's CUDA-X domain-specific libraries offer optimized implementations of complex algorithms for these automated agents. Researchers using agent swarms for large-scale problem-solving could see a substantial performance boost from this hardware and software combination.
Comparison and context
When placed next to consumer graphics hardware, the scale of the Rubin HPC becomes clear. It offers nearly 2.5 times as many CUDA cores as Nvidia's flagship GeForce RTX 5090, which is built for gaming and creative work.
| Component | CUDA Cores | Primary Optimization |
|---|---|---|
| Nvidia Rubin HPC GPU | >50,000 | Scientific simulation, data analytics |
| Nvidia GeForce RTX 5090 | 21,760 | Gaming, content creation |
Unlike other high-end GPUs focused on AI training and inference, the Rubin HPC is specifically optimized for scientific simulations and data analytics. Scientific computing workloads typically involve applications accessing immense datasets, gathering scattered values, calculating results, and repeating the process at an enormous scale. The Rubin HPC's delayed release in 2028 positions it roughly two years after the standard Rubin generation, making it a specialized successor for the most demanding computational science tasks. Ian Buck, vice president of hyperscale and high-performance computing at NVIDIA, led the announcement.





