Nvidia가 NVLink 7과 5만 개 이상의 CUDA 코어를 탑재한 Rubin HPC GPU를 공개했다.
Rubin HPC is a new GPU in the NVIDIA Vera Rubin platform. Image: NVIDIA
NVIDIA is packing a large number of CUDA cores into its next-generation Rubin HPC GPU, a new addition to its Vera Rubin platform. Announced at the SC26 supercomputing conference, the Rubin HPC GPU is a high-end chip with more than 50,000 CUDA cores on a single processor. That is a record for NVIDIA, though core count is only part of the story.
The other notable highlight is that the new accelerator introduces seventh-generation NVLink connectivity to keep all those cores fed with data. Beyond some previous roadmap teasers, NVIDIA has not disclosed many details about NVLink 7, but for context, NVLink 6 provides 3TB/s of bandwidth per GPU.
While Rubin HPC joins NVIDIA's broader Vera Rubin platform, it is tailored for high-performance computing (HPC) rather than focused primarily on training and running AI models. It targets workloads such as complex scientific simulations and engineering calculations, where moving massive datasets efficiently can be just as important as processing them.
There is an AI angle as well. NVIDIA envisions a future in which AI agents help scientists and engineers generate and evaluate new ideas, relying on GPU-powered simulations to determine which ones are worth pursuing.
"For agents tackling science and engineering, faster simulation means faster discovery," said Ian Buck, vice president of hyperscale and high-performance computing at NVIDIA. "Rubin HPC accelerates the simulations agents use to test ideas and guide their next steps, helping scientists and engineers explore more possibilities and bring breakthroughs closer."
NVIDIA Rubin HPC rack and GPU compute tray. Image: NVIDIA
The core count is what really sets Rubin HPC apart. For perspective, NVIDIA's standard Vera Rubin GPU packs 28,672 CUDA cores and 224 streaming multiprocessors, which itself dwarfs the consumer-grade GeForce RTX 5090 with its 21,760 CUDA cores.
NVIDIA has yet to reveal several important specifications, including memory capacity, bandwidth, and power consumption, in addition to the details of NVLink 7. Those details will come at a later date.
As for availability, NVIDIA says the Rubin HPC GPU is expected to arrive in 2028.