News 3 min read

NVIDIA Rubin HPC: a GPU with over 50,000 CUDA cores, due in 2028


NVIDIA announced Rubin HPC on November 16, 2026, at the SC26 supercomputing conference: a new GPU in its Vera Rubin family built for simulation and data processing in science, engineering and analytics. It packs more than 50,000 CUDA cores on one chip, which NVIDIA calls a record, and adds hardware to keep sparse data flowing to those cores. NVIDIA expects it to be available in 2028, so it is a roadmap item, not something you can rent this year.

What’s new

From NVIDIA’s Rubin HPC announcement:

  • More than 50,000 CUDA cores on a single GPU.
  • Sparse data acceleration, a feature meant to keep those cores fed with data in workloads where most values are zero or scattered across memory.
  • Seventh-generation NVLink to connect GPUs to each other.
  • A “balanced design” that, in NVIDIA’s words, scales data movement and compute together.
  • Target workloads: sparse matrix operations, stencil computations, sorting and reductions, the building blocks NVIDIA says recur across climate models, engineering simulation, genomics and financial risk.

NVIDIA pitches it as hardware for AI agents that run scientific software. “For agents tackling science and engineering, faster simulation means faster discovery,” said Ian Buck, NVIDIA’s vice president of hyperscale and high-performance computing.

The announcement gives no figures for FP64 performance, memory capacity, memory bandwidth or power, and names no customers or systems yet.

What it means for developers

Nothing changes in your code today. Rubin HPC is a CUDA GPU, so software written against CUDA and NVIDIA’s CUDA-X libraries is the path onto it; NVIDIA says those libraries are where it puts optimized versions of complex algorithms for agents, especially for simulation and analytics. If you maintain a simulation code, the workloads NVIDIA lists (sparse linear algebra, stencils, sorts, reductions) are the ones it expects to speed up, so they are the parts worth profiling now.

At the same event NVIDIA published a round-up of new CUDA-X libraries for science, which run on current GPUs. Two are newly open source: cuSpa, for cell polygon calculations, which NVIDIA says ran 900x faster than on CPUs at Memorial Sloan Kettering Cancer Center, and a GPU-accelerated HMMER, which ran benchmark searches up to 84x faster than standard HMMER for Basecamp Research. Another, cuPotential, is coming to molecular dynamics tools such as GROMACS, LAMMPS, NAMD and AMBER, with up to 3x faster simulation claimed.

NVIDIA’s agent tooling for these workloads is also open: we covered its donation of the OpenShell agent sandbox runtime to the CNCF last week.

How it compares

Rubin HPC sits next to the existing Vera Rubin parts, which are aimed at both AI and science. In June, at ISC 2026, NVIDIA said a Vera Rubin rack with up to 144 GPUs delivers 5 petaflops of native FP64 and more than 7 exaflops of AI for science, and that Vera Rubin NVL4 systems would ship from Bull, Dell, GIGABYTE, HPE and Supermicro in the fourth quarter of 2026. Supercomputers announced on that platform include Blue Lion at the Leibniz Supercomputing Centre, due online in 2027, and Doudna at NERSC. Rubin HPC is a separate, simulation-focused chip due in 2028, after those systems, and NVIDIA has not yet said how its performance compares with them.

Sources