NVIDIA Plans 576-GPU Rubin Ultra Racks for 10 Trillion-Parameter AI Models

NVIDIA plans to connect 576 Rubin Ultra GPUs in a rack-scale system designed to support AI models with 10 trillion parameters, according to Wccftech.

The report says the chips will use next-generation co-packaged optics (CPO) and NVLink to operate as a single GPU. It describes Rubin Ultra as the platform following Rubin, and says it will use NVIDIA’s Kyber rack, a fully liquid-cooled design.

The article contrasts Kyber with Vera Rubin “VR200” platforms, which it says are based on Oberon NVL72 servers. It attributes the rack specifications to details NVIDIA unveiled at GTC, but the supplied source excerpt does not include the full specifications.

Wccftech says the system is intended to deliver the “highest performance per watt” on the planet. That is the report’s characterization; the provided material does not include comparative measurements to substantiate it.