NVIDIA announces Rubin HPC GPU for science and engineering
NVIDIA introduced Rubin HPC, a scientific-computing GPU for simulation and data-intensive workloads that the company expects to make available in 2028.
NVIDIA announced Rubin HPC at SC26 on Monday, adding a scientific-computing GPU to its Vera Rubin platform that the company expects to make available in 2028. The planned chip targets simulation and data processing across science, engineering and analytics rather than the agentic AI focus of the previously announced Rubin GPU.
NVIDIA said Rubin HPC will contain more than 50,000 CUDA cores and use sparse-data acceleration to keep those computing cores supplied with useful data when workloads contain many empty or zero values. The company also plans to use seventh-generation NVLink for connections between GPUs.
The design is intended to scale computation and data movement together for recurring scientific operations, including sparse matrix calculations, stencil computations, sorting and reductions. NVIDIA has not provided measured Rubin HPC benchmark results, and its claims that the chip will shorten simulations and discovery or design cycles remain forecasts.
Rubin HPC is separate from the Rubin GPU architecture, which NVIDIA describes as having 224 streaming multiprocessors, 896 Tensor Cores, HBM4 memory and sixth-generation NVLink. It is also distinct from the Vera Rubin system lineup, including NVL72, DGX Rubin NVL8, HGX Rubin NVL8 and NVL4 configurations.
The timing extends beyond the initial Rubin rollout. NVIDIA previously said Rubin-based products would be available from partners in the second half of 2026, while Rubin HPC is expected in 2028. The company has not disclosed the new GPU’s price, memory capacity, power envelope, system configuration, launch customers or a more specific availability date.
More news

Oracle and NVIDIA make Vera Rubin NVL72 and Vera CPUs available on OCI

NVIDIA partners set out European AI capacity plans

NVIDIA says it contributed Vera Rubin rack designs to OCP
