From the source
Nvidia introduced Rubin HPC, an upcoming graphics card with over 50,000 CUDA cores, alongside more than a half-dozen scientific software tools at the SC26 supercomputing conference in Chicago.
- The chip is aimed at scientific simulations as well as large language models, and Nvidia says it will ship in 2028, so customers must rely on the software tools until then.
- Rubin HPC adds a "sparse data acceleration" feature for datasets full of zeros, which Nvidia says are common in biology and astrophysics, and it ships with a new iteration of NVLink, whose current version moves 3.6 terabits per second per card.
- Among the software additions, cuPhoton.xBeam analyzes Event Horizon Telescope data and, according to Nvidia, performs analysis that historically took a month in eight minutes.
- Nvidia is open-sourcing cuSpa for cell polygon calculations and packaging the existing rapids-singlecell tool in a container, while Nemotron for Science bundles the open-source Nemotron 3.5 Super VL model with post-training workflows and nearly 50,000 chemistry questions.
- Nvidia also expanded CUDA-Q with quantum software building blocks and a qubit error mitigation algorithm, and is working with IBM and Fermilab on a modular quantum codesign platform combining DGX B200 systems with IBM quantum hardware.