Nvidia unveils Rubin HPC with more than 50,000 CUDA cores
The Vera Rubin GPU is meant to speed the simulations that software agents run for science, engineering and analytics, with sparse-data hardware and seventh-generation NVLink. Nvidia withheld memory size, memory speed and any FP64 mix.
Nvidia announced Rubin HPC at Supercomputing 2026, a graphics processor in the Vera Rubin platform for scientists, engineers and analysts. Software agents can try more molecules, materials and designs; the chip is meant to run the simulations that test those ideas and point to the next ones.
Nvidia says a single Rubin HPC GPU holds more than 50,000 CUDA cores, which it calls a record. The Rubin GPU already on sale has 28,672 CUDA cores, Tom's Hardware reported. Hardware aimed at sparse data is supposed to keep that larger core count supplied, and links between GPUs use the seventh generation of NVLink. Nvidia describes the part as balanced, with data movement scaled alongside compute so the cores spend time calculating. It highlights recurring work such as reductions, sorting, stencil calculations and sparse matrix operations, the patterns it ties to financial-risk models, genomic analysis, engineering simulation and climate models.
Figures Nvidia did not give
Nvidia published no HBM capacity or bandwidth for Rubin HPC, and it did not say whether compute is shifted toward native FP64, the approach AMD and Intel have been promoting with their own GPUs at the conference. It also gave no fresh NVLink bandwidth number inside a rack. Earlier descriptions of that interconnect indicated per-GPU bandwidth stays at 3.6 TB/s, the NVLink 6 rate, while the scale-up domain grows.
Ian Buck, vice president of hyperscale and high-performance computing at Nvidia, said that for these agents “faster simulation means faster discovery.”