NVIDIA Reveals Rubin Ultra NVL576 Architecture: Connecting 576 GPUs into One Domain via CPO
NVIDIA disclosed its next-gen AI infrastructure blueprint at OCP, using co-packaged optics to link 576 GPUs, aiming to support trillion-parameter models and boost performance per watt.
US chipmaker NVIDIA further revealed plans for its Rubin Ultra platform at the Open Compute Project (OCP) 2026 conference on October 22. The company announced it will use next-generation Co-Packaged Optics (CPO) technology and NVLink 7 to connect 576 GPUs into a single compute domain. The goal is to make an entire liquid-cooled rack cluster operate like one giant GPU, claiming the "highest performance per watt" globally.
The Vera Rubin Ultra NVL576 architecture is designed specifically for ultra-large AI models with over 10 trillion parameters. According to previously published specifications, the platform's target performance reaches 15 exaflops for FP4 inference and 5 exaflops for FP8 training, approximately 14 times higher than the Blackwell GB300 NVL72. In terms of memory, the entire platform can be configured with up to 365TB of HBM and 4.6PB/s of total memory bandwidth.
NVIDIA also showcased the new NVLink 7 CPO Switch Tray and Vera Rubin Ultra Compute Tray. The former extends NVLink connectivity optically, while the latter integrates two Rubin Ultra GPUs and one Vera CPU. The company emphasized that the overall architecture strictly follows OCP standards, aiming to reduce installation time from hours to minutes.
It should be noted that these specifications remain technology roadmap targets, and the final commercial version may vary. This disclosure indicates NVIDIA is pushing rack-scale systems toward higher density, higher bandwidth, and greater reliance on optical interconnect technologies.