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 - Wccftech

NVIDIA's upcoming Rubin Ultra GPU platform will represent a significant leap in computing power, featuring a massive 576-GPU rack-scale solution. All GPUs will be connected using next-generation Co-Packaged Optics (CPO) and NVLink technologies, effectively creating one giant GPU with the highest performance-per-watt ratio on the planet.

The Rubin Ultra NVL576 will offer 15 Exaflops of FP4 inference performance and 5 Exaflops of FP8 training performance, representing a 14x improvement over the current Blackwell GB300 NVL72. This massive computational boost will enable the system to handle AI models with 10 trillion or more parameters.

Memory and Connectivity

Each Rubin Ultra GPU will feature up to 1 TB of HBM4E memory across a 4-reticle-sized GPU, scaling memory capacity by 8x compared to previous generations. The entire rack will deliver 365 TB of HBM memory with an impressive 4.6 PB/s bandwidth. Additionally, the platform will include a next-generation 1.5 PB/s NVLink7 interconnect and a new ConnectX-9 Spectrum chip with 115.2 TB/s throughput.

Design and Deployment

The Vera Rubin Ultra Compute Tray will pack two Rubin Ultra GPUs and a single Vera CPU, all liquid-cooled for optimal thermal management. NVIDIA has also unveiled its NVLink 7 CPO Switch Tray for the first time, which will be featured in each rack. The design is based on Open Compute Project (OCP) standards, featuring modular components that reduce installation time from hours to minutes.

NVIDIA's CEO Jensen Huang recently confirmed during GTC Berlin that CPO racks are now in volume production, addressing previous reports of delays. The Rubin Ultra will be the first rack-level system to fully utilize this technology, which is expected to become the industry standard for high-performance computing.

Industry Impact

The Rubin Ultra platform marks NVIDIA's decisive move toward treating an entire 576-GPU liquid-cooled rack as one coherent superchip. By combining massive scale, extreme efficiency, and OCP-based modular design, NVIDIA aims to establish CPO as the new interconnect standard while providing the performance-per-watt foundation needed for next-generation AI models.

To facilitate deployment, NVIDIA has developed reference designs like the DSX Pod, which reduces deployment time from months to just 14 days. This acceleration will be crucial for organizations looking to implement massive AI infrastructure while managing capital expenditures effectively.