Nvidia announced a new GPU called Rubin HPC at the SC26 supercomputing conference, the company said in a blog post. It belongs to the Vera Rubin platform. Nvidia says it has more than 50,000 CUDA cores, the small processing units that do a GPU's math, on a single chip. It also adds a feature for sparse data, meaning data that is mostly empty, and uses seventh-generation NVLink to link GPUs together. Nvidia says Rubin HPC is meant to speed up simulation and data work for science, engineering and analytics, including work run by AI agents. It expects the chip in 2028.
Tom's Hardware reported that today's Rubin GPU has 224 streaming multiprocessors and 28,672 CUDA cores. It said a jump to more than 50,000 would probably need more chips on one package. That is the outlet's inference, not something Nvidia stated. Nvidia gave no details on the chip's layout, memory size or memory speed, Tom's Hardware said. Nvidia also did not say whether it shifts more of the chip toward 64-bit math, which many scientific codes need. Tom's Hardware noted that AMD and Intel are promoting that this week. Nvidia gave no pricing and no exact release date.
For teams that plan or buy compute for simulation, analytics or agent-driven research, there is now a stated year, 2028. The facts that decide fit and cost are still missing. Those are memory capacity, memory speed and double-precision performance.