DEV Community

Cover image for Nvidia unveils Rubin HPC GPU with more than 50,000 CUDA cores
The daily flare
The daily flare

Posted on Originally published at thedailyflare.com

Nvidia unveils Rubin HPC GPU with more than 50,000 CUDA cores

Originally published on The Daily Flare.

Nvidia announced Rubin HPC, a new GPU for scientific computing, at the SC26 conference in Chicago on November 16, saying it has more than 50,000 CUDA cores and is due in 2028.

The chip is part of Nvidia’s Vera Rubin platform and is designed for science, engineering and analytics workloads. In its company announcement, Nvidia said the GPU combines its record core count with sparse-data acceleration and seventh-generation NVLink connectivity. The company says the combination is intended to keep data moving to the GPU’s compute resources as simulations process large datasets.

The core count marks a substantial increase over the current Rubin GPU. Tom’s Hardware reported that the current model has 224 streaming multiprocessors and 28,672 CUDA cores. Nvidia has not disclosed Rubin HPC’s physical layout. Tom’s Hardware said the greater-than-50,000 figure could imply a multi-die design similar to Rubin Ultra, which links four dies using NVLink-C2C, but described that as a possibility rather than a confirmed configuration.

Built for data-intensive simulations

Nvidia describes Rubin HPC as a balanced design that scales data movement alongside computing capacity. The company says this is important for scientific workloads in which applications repeatedly retrieve scattered values from large datasets, perform calculations and repeat the process at scale. It identifies sparse matrix operations, stencil computations, sorting and reductions as recurring patterns the GPU is intended to accelerate.

The workloads Nvidia highlighted include climate modelling, engineering simulations, genomic analysis and financial risk calculations. The company says these applications can involve large datasets and repeated simulation runs. Nvidia argues that improving both computation and data movement can reduce the time GPU resources spend waiting for information, allowing teams to complete simulations more quickly.

The announcement is tied to Nvidia’s push to use AI agents in research and engineering. The company says agents can explore candidate materials, molecules and designs, then use simulations to test ideas and guide what to investigate next. Ian Buck, Nvidia’s vice president of hyperscale and high-performance computing, said: “For agents tackling science and engineering, faster simulation means faster discovery.”

Nvidia also pointed to its CUDA-X libraries, which provide optimized implementations of algorithms for simulation and analytics. The company says agents can use those libraries with Rubin HPC to run simulations and receive results sooner, shortening the cycle between forming a hypothesis, testing it and refining the next step. The intended applications include work by scientists, engineers and analysts handling large datasets.

At a November 17 presentation at SC26, Nvidia said Rubin HPC supports precisions ranging from FP64 to FP4. The company’s announcement also describes sparse-data acceleration as a way to keep its CUDA cores supplied with information. Nvidia has not announced additional specifications for the GPU’s memory capacity or bandwidth in the material about the launch.

For communication between GPUs, Rubin HPC will use NVIDIA NVLink 7. Nvidia says the seventh-generation interconnect will accelerate GPU-to-GPU connectivity. Tom’s Hardware noted that the announcement did not give new figures for NVLink 7’s bandwidth or describe its physical implementation within a rack.

Nvidia expects Rubin HPC to become available in 2028. The company presented it as a GPU for scientific computing and agent-driven workflows, while the chip’s detailed physical configuration and further performance specifications remain unannounced.

Top comments (0)