DGX SuperPOD

DGX SuperPOD is NVIDIA's turnkey reference architecture for clusters of DGX systems, sold as one product with a specific bill of materials, installation, and support. It grows in scalable units, or SUs, each a defined block of DGX systems and the leaf switches that serve them. Separate networks carry compute, storage, in-band management, and out-of-band management traffic. NVIDIA Mission Control software administers it.

SU size follows the DGX generation. One SU holds 32 DGX H100 systems, 64 DGX B300 systems, or eight DGX Vera Rubin NVL72 racks. Those eight racks carry 576 GPUs and need about 2 MW of thermal design power. The DGX B300 design runs its compute fabric at 800 Gb/s over XDR InfiniBand or Spectrum-X Ethernet. DGX BasePOD is the more flexible sibling. It contains DGX systems, NVIDIA networking, NVIDIA software including Base Command Manager, and certified third-party storage, and storage vendors may customize it.

Choose SuperPOD when the cluster will follow NVIDIA's bill of materials from one SU up. Choose BasePOD when the plan needs a different cluster manager or job scheduler, storage outside the SuperPOD certified list, or a changed fabric topology. NVIDIA says those changes generally disqualify a SuperPOD.

Sources

Last verified October 7, 2026.