Nvidia is using its DGX SuperPOD as the reference architecture for large scale systems built on its newly announced Rubin platform, which the company describes as its next major leap in artificial intelligence computing. Unveiled at CES in Las Vegas, the Rubin platform combines six new chips into a tightly codesigned artificial intelligence supercomputing stack aimed at accelerating agentic artificial intelligence, mixture of experts models, and long context reasoning, while also reducing the cost of inference token generation. The platform integrates the Nvidia Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, and Spectrum-6 Ethernet Switch into a single system design that spans compute, networking, and software.
Nvidia highlights five key advances in the Rubin platform used in new DGX systems, including sixth generation Nvidia NVLink, which delivers 3.6TB/s per GPU and 260TB/s per Vera Rubin NVL72 rack for mixture of experts and long context workloads. The Rubin GPU provides 50 petaflops of NVFP4 compute for artificial intelligence inference and incorporates a third generation Transformer Engine with hardware accelerated compression, while the Vera CPU offers 88 custom Olympus cores with Armv9.2 compatibility and ultrafast NVLink-C2C connectivity. Third generation Nvidia Confidential Computing is supported at rack scale through Vera Rubin NVL72, and a second generation RAS engine extends health monitoring and fault tolerance across GPU, CPU, and NVLink, with tray designs that enable 3x faster servicing. Nvidia states that together these innovations deliver up to 10x reduction in inference token cost of the previous generation as artificial intelligence models expand in size and reasoning depth.
The DGX SuperPOD architecture is being extended with Rubin based systems through two primary configurations. Nvidia DGX SuperPOD with DGX Vera Rubin NVL72 aggregates eight DGX Vera Rubin NVL72 systems, featuring 576 Rubin GPUs, to provide 28.8 exaflops of FP4 performance and 600TB of fast memory, with 260TB/s of aggregate NVLink throughput allowing the entire rack to function as a single artificial intelligence engine and eliminating the need for model partitioning. Each DGX Vera Rubin NVL72 combines 36 Vera CPUs, 72 Rubin GPUs, and 18 BlueField-4 DPUs to create a unified memory and compute space, while DGX SuperPOD with DGX Rubin NVL8 scales out 64 DGX Rubin NVL8 systems featuring 512 Rubin GPUs in a liquid cooled, x86 based form factor that Nvidia says delivers 5.5x NVFP4 FLOPS compared with Nvidia Blackwell systems. Across configurations, DGX SuperPOD integrates BlueField-4 DPUs, ConnectX-9 SuperNICs, Nvidia Quantum-X800 InfiniBand, Nvidia Spectrum-X Ethernet, and Nvidia Mission Control to provide secure, software defined infrastructure and automated artificial intelligence operations.
Nvidia positions Rubin based DGX SuperPOD as the networking and orchestration backbone for what it calls gigawatt artificial intelligence factories. The platform includes an 800Gb/s end to end networking stack with Nvidia Quantum-X800 InfiniBand, which uses SHARP v4 and adaptive routing to offload collective operations, and Nvidia Spectrum-X Ethernet, built on Spectrum-6 switches and ConnectX-9 SuperNICs to optimize east west artificial intelligence traffic patterns over standard Ethernet. On the software side, Nvidia Mission Control, originally designed for Blackwell based DGX systems, will manage Rubin based DGX deployments, handling configuration, facilities integration, cluster and workload management, as well as power, cooling, and resiliency functions such as rapid leak detection and autonomous recovery. DGX systems also support the Nvidia Artificial Intelligence Enterprise platform and Nvidia NIM microservices, including for the Nemotron-3 family of open models. Nvidia says DGX SuperPOD combined with the Rubin platform will serve as the launchpad for a new generation of industrial artificial intelligence, enabling systems that can reason across thousands of steps and support frontier models and agentic artificial intelligence applications at significantly lower cost, with Rubin based DGX SuperPOD systems planned for availability in the second half of this year.
