What Is DGX SuperPOD?_
The DGX SuperPOD is a reference architecture that bundles multiple DGX base pods (typically 8 DGX nodes each) with high-speed networking and storage to form a unified supercomputer. It uses NVIDIA's NVLink fabric for GPU-to-GPU communication within each node and InfiniBand fabric for communication between nodes across the supercomputer. The system is factory-integrated and validated to reduce deployment risk and time for enterprise data centers.
Technical Details
Each DGX node in a SuperPOD uses NVLink (via NVSwitch) for GPU-to-GPU communication within the node, while InfiniBand (typically HDR or NDR) is used for scale-out networking between nodes. The architecture includes a dedicated storage tier (e.g., DDN or NetApp) and a management network for orchestration. Cooling can be air or liquid, depending on the DGX model (e.g., H100), with power distribution requiring high-density PDUs and redundant feeds.
How Leviathan Systems Works with DGX SuperPOD
In our field work, deploying a DGX SuperPOD involves racking DGX nodes in pre-defined base pod layouts, running structured copper cabling for NVLink and fiber cabling for InfiniBand per the NVIDIA cabling guide, and commissioning the InfiniBand fabric with the Leviathan team's network engineers. We also handle liquid cooling connections for newer DGX models and verify NVSwitch seating during assembly.
Related service
GPU Rack Assembly & Rack-and-Stack →