NVIDIA details 576-GPU Rubin Ultra NVL576 CPO platform
At OCP 2026, NVIDIA detailed a roadmap for connecting 576 Rubin Ultra GPUs across racks through next-generation CPO and NVLink 7 as one compute domain. The liquid-cooled Vera Rubin Ultra NVL576 design targets AI models with more than 10 trillion parameters, with vendor-stated goals of 15 exaflops of FP4 inference and 5 exaflops of FP8 training performance. NVIDIA also showed the NVLink 7 CPO Switch Tray and a Compute Tray containing two Rubin Ultra GPUs and one Vera CPU.
The disclosed rack-scale architecture would extend NVIDIA's AI infrastructure roadmap toward very large model workloads, but its performance, efficiency and installation-time claims remain unverified vendor targets.
NVIDIA To Connect 576 Rubin Ultra GPUs Using CPO To Create One Giant GPU With The “Highest Performance per Watt” on the Planet Next Year To Drive 10 Trillion-Parameter AI Models
Hassan Mujtaba•Oct 22, 2026 at 11:35am EDT
Add to Preferres Sources
NVIDIA's Rubin Ultra GPUs will come in a massive 576-GPU rack-scale solution, all connected using next-gen CPO and NVLINK technologies.
Some source content is omitted. Continue reading on the original site.
#NVIDIA#NVIDIA Rubin Ultra NVL576#NVLink 7#NVLink 7 CPO Switch Tray#Vera Rubin Ultra Compute Tray#Vera CPU#Kyber rack#HBM4E#Blackwell GB300 NVL72