Nvidia adds a GPU for scientific computing to its Rubin platform

Illustrative image: Rafael Minguet Delgado on Pexels
Nvidia has announced Rubin HPC, a scientific computing GPU with more than 50,000 CUDA cores and dedicated hardware for sparse data processing. Expected in 2028 as part of the Vera Rubin platform, it targets simulation workloads in which locating and moving data can occupy as much of the calculation as the arithmetic itself.
Sparse matrices arise when a model records only the interactions that exist between its elements, leaving most potential entries at zero. Storing the nonzero values saves memory, but the processor then needs indices to locate them and the corresponding inputs. The resulting accesses can be scattered, leaving arithmetic units waiting for data even when considerable computing capacity is available.
Compressed sparse row storage, documented in Nvidia’s cuSPARSE library, illustrates this pattern. One array holds the nonzero values, another their column indices and a third marks the start of each row. A multiplication can read sequentially through the stored values while fetching widely separated entries from the vector they multiply, combining an orderly traversal with irregular memory accesses.
Rubin HPC’s announced sparse data hardware addresses this class of workload, alongside acceleration Nvidia describes for stencil calculations, sorting and reductions. These operations differ in how they reuse values and move through memory: a stencil works with neighbouring points, while a reduction combines many values into a smaller result. Nvidia has not yet detailed the new acceleration circuitry or published comparative benchmarks, memory capacity or power specifications.
Simulations spread across multiple GPUs add exchanges between processors to the movement within each device. Rubin HPC will use Nvidia’s seventh generation of NVLink for those connections, linking the work on local data to the wider system’s communication requirements. The processor was announced at SC26, with the first shipping systems and their configurations still to be identified ahead of expected availability in 2028.
Sources: Nvidia’s Rubin HPC announcement; Nvidia cuSPARSE documentation on compressed sparse row storage.



