NVIDIA targets scientific simulations with Rubin HPC, due in 2028
NVIDIA announced a Rubin HPC GPU with more than 50,000 CUDA cores, sparse data acceleration and seventh-generation NVLink. The design targets simulations used by scientists and AI agents, but missing performance, power and pricing details leave its practical advantage unproven.
- NVIDIA expects Rubin HPC in 2028, targeting scientific simulations and analytics.
- Sparse data acceleration and NVLink 7 aim to keep calculations supplied with data.
- Undisclosed performance, power and pricing prevent a meaningful assessment of value.
NVIDIA, the GPU and computing-platform supplier whose earlier H100 processor was designed for both AI and scientific computing, announced Rubin HPC on November 16 at the SC26 supercomputing conference. The company promises more than 50,000 CUDA cores on one GPU and expects availability in 2028 as part of its Vera Rubin computing platform.
In its announcement, NVIDIA argues that AI agents—software systems that propose ideas and invoke simulations to test them—will increase demand for scientific computation. Rubin HPC is meant to help researchers evaluate more candidate materials, molecules and designs. Its proposed advantage depends on moving scattered data efficiently as well as adding arithmetic capacity; the announcement supplies no application benchmarks establishing the gain.

Rack and tray rendering published with NVIDIA’s Rubin HPC announcement. Image: NVIDIA. Source: NVIDIA.
Keeping scientific calculations supplied with data
Rubin HPC is a specialized addition to a platform whose regular Rubin GPU emphasizes AI workloads. HPC means high-performance computing; here, the target applications include climate modeling, engineering simulations, genomic analysis and financial risk calculations.
NVIDIA describes a recurring problem in those applications: calculations retrieve scattered values from enormous datasets, perform an operation and repeat. Adding arithmetic capacity alone offers limited help if that capacity spends its time waiting for data.
The company says Rubin HPC will address that problem with sparse data acceleration, alongside what it calls a record number of CUDA cores, the GPU’s arithmetic processing units. Its stated targets include sparse matrix operations, stencil computations, sorting and reductions. These cover tasks such as working with matrices containing many zero entries and updating values using neighboring points in a simulation grid.
The proposed mechanism remains only partly described. A technical analysis of the announcement notes that NVIDIA has not explained whether the sparse data capability relies on additional functional units or particular memory structures. The same analysis interprets NVIDIA’s simplified diagram as suggesting at least two silicon dies combined into one GPU, as with regular Rubin. That is an interpretation of the diagram, not a confirmed Rubin HPC specification: “one GPU” should not be read as a claim about a single die.

