NVIDIA's Rubin Ultra GPUs will form a 576-GPU rack-scale solution connected via next-gen CPO and NVLINK technologies. The system will operate as a single unit, offering 15 Exaflops of FP4 inference and 5 Exaflops of FP8 training performance, a 14x improvement over Blackwell GB300 NVL72.
Ian Buck stated that all 576 GPUs in the Kyber Rack will use NVLink to function as one unit, claiming it will be the fastest GPU with the highest performance per watt. The design includes liquid cooling, up to 1 TB of HBM4E memory per GPU, and OCP-standard modular components for simplified installation.