Nvidia on November 16, 2026 announced Rubin HPC, a new GPU in its Vera Rubin platform designed to accelerate simulation and data processing for science, engineering and analytics. The company unveiled the chip at SC26, the annual supercomputing conference, and says it is expected to be available in 2028.
Nvidia's pitch ties the chip to AI agents. In its announcement, the company argues that as agents help researchers and engineers explore more candidate materials, molecules and designs, demand grows for the simulations used to test those ideas. Rubin HPC is meant to speed up those simulations and shorten the loop of hypothesis, simulation and refinement.
The headline specification is core count. Nvidia says Rubin HPC has more than 50,000 CUDA cores on a single GPU, which it calls a record. Tom's Hardware notes that today's Rubin GPU has 28,672 CUDA cores across 224 streaming multiprocessors, and suggests that reaching the new figure may require a multi-die package similar to Rubin Ultra, though Nvidia did not disclose the chip's physical layout. HPCwire describes the chip as having three times the vector cores of Blackwell.
Nvidia frames the design as balanced rather than compute-only. Scientific workloads such as climate models, engineering simulations, genomic analysis and financial risk tend to gather scattered values from huge datasets, so the GPU adds what Nvidia calls a sparse data acceleration capability to keep its cores supplied with data. The company says this targets recurring patterns including sparse matrix operations, stencil computations, sorting and reductions. GPU-to-GPU traffic runs over seventh-generation NVLink, and Nvidia points to its CUDA-X libraries as the software agents would call.
HPCwire reads the chip as a renewed commitment to FP64, the double-precision math traditional supercomputing relies on, after Nvidia's recent data center generations emphasized lower-precision AI formats. Nvidia's Dion Harris told the publication that customers may need a GPU suited to running simulations across a broad range of precisions.
Key details are still missing. Tech Times and Tom's Hardware both note that Nvidia has not published memory capacity or bandwidth, FP64 throughput, power draw or pricing, and the announcement names no launch customers. That matters for the sparse, memory-bound codes Nvidia is targeting, where data movement rather than raw arithmetic often limits real-world performance.
For now, Rubin HPC is a roadmap signal: Nvidia is positioning a dedicated simulation GPU alongside its AI accelerators on the expectation that agent-driven research will multiply the number of simulations scientists and engineers need to run. Performance claims will need independent testing once hardware and full specifications arrive.
