NVIDIA unveils Rubin HPC GPU with over 50,000 CUDA cores, expected in 2028
Edited and reviewed by AI Ping Editorial Team
3-line summary
- NVIDIA unveiled Rubin HPC at the SC26 conference in Chicago for scientific and engineering simulations.
- It says each GPU will have more than 50,000 CUDA cores, sparse-data acceleration and seventh-generation NVLink.
- NVIDIA expects the GPU to become available in 2028. It did not disclose memory specifications, pricing or measured runtimes.
NVIDIA unveiled Rubin HPC, a GPU for scientific computing, at the SC26 supercomputing conference in Chicago on November 16. It says the GPU will have more than 50,000 CUDA cores, which perform calculations in parallel, and expects it to become available in 2028. The chip is intended for work such as simulating candidate designs that researchers and engineers generate with AI.
Rubin HPC is a new member of the Vera Rubin platform NVIDIA introduced at CES in January. The NVL72 name on an existing product refers to a rack with 72 Rubin GPUs and 36 Vera CPUs. The figure of more than 50,000 describes the number of CUDA cores on one Rubin HPC GPU. It gives a count of computing resources. Runtime for an individual job still depends on the software and the conditions under which it runs.
Computing the tests that follow a proposed design
In the workflow NVIDIA outlined, AI agents help researchers explore a larger pool of materials or designs. Before committing to physical prototypes or laboratory tests, researchers can feed the candidates into a numerical simulation, compare the results and decide what to test next. The new GPU is designed for the computations and data handling that happen during the simulation. Ian Buck, NVIDIA’s head of hyperscale and high-performance computing, said the aim is to shorten that stage so teams can choose their next experiments sooner.
One example is computational fluid dynamics, which a design team can use to see how changes in shape affect airflow before building a prototype. NVIDIA’s existing guidance describes using Blackwell GPUs to run AI-based predictions alongside detailed numerical solvers. The guidance explains how a quick estimate can be followed by a higher-fidelity simulation to validate it.
Keeping data moving within and across GPUs
A weather model or engineering simulation may repeatedly pick values from scattered locations in a large data set. A sparse matrix, in which most entries are zero, can be represented by storing the entries that have values and where they occur. NVIDIA’s existing sparse-computing documentation describes formats that do this. NVIDIA says Rubin HPC will include acceleration for working with sparse data.
When a calculation is shared across GPUs, how quickly data moves between devices also affects the time to finish. NVIDIA says Rubin HPC will use its seventh-generation NVLink interconnect. The earlier Vera Rubin NVL72 rack uses the sixth generation. The new announcement identifies the interconnect but gives no aggregate data transfer rate for a Rubin HPC system. It cites operations such as stencils, which use neighboring values on a grid, and sorting or combining many entries. In each case, performance depends on getting the relevant inputs to the cores as well as carrying out the calculation.
What an existing Rubin trial measured
One example in NVIDIA’s scientific software lineup is cuPhoton, which aligns and correlates signals from radio telescopes. In a test NVIDIA published on November 13, 64 GPUs in an existing Vera Rubin NVL72 system took 70 seconds to read and correlate 10 terabytes of observations. The software divided observation time into chunks processed by different GPUs and combined the results. Correlation prepares inputs for image reconstruction. NVIDIA ran the test on the existing rack, and the Rubin HPC announcement gives no measurement of the new GPU doing the same job.
At SC26, NVIDIA separately introduced a scientific-computing rack with 192 Vera CPUs. It expects the CPU rack to become available through Dell, HPE, Lenovo and Supermicro beginning in early 2027. The German Climate Computing Center says the supercomputer it expects to put into service in early 2028 will have a Vera CPU partition and a separate GPU partition built on GB200 Grace Blackwell for most of the center’s simulation and AI computing.
NVIDIA expects to make the Rubin HPC GPU available in 2028. NVIDIA has yet to give a more precise date or specify its price, memory capacity or memory bandwidth.
Sources
- Rubin HPC announcement and expected availability — NVIDIA
- SC26 event schedule and location — NVIDIA
- January Vera Rubin platform introduction — NVIDIA Korea
- Computational fluid dynamics simulation — NVIDIA
- cuSPARSE storage formats — NVIDIA Documentation
- CUDA-X libraries for scientific computing — NVIDIA
- cuPhoton experiment on Vera Rubin NVL72 — NVIDIA Technical Blog
- Vera CPU rack announcement and expected schedule — NVIDIA Newsroom
- HLRE-5 climate supercomputer configuration — German Climate Computing Center
- Existing Vera Rubin NVL72 rack specification — NVIDIA Newsroom