Separate compute and storage fabric decisions in an XDR AI cluster

NVIDIA’s DGX B300 SuperPOD reference architecture provides a concrete design example: its major-components table lists Q3400-RA order code 920-9B36F-00RX-8S0 for compute leaves and spines, and SN5610 order code 920-9N42F-00RI-3C1 for Ethernet storage leaves and spines. These are different network roles, even though both platforms advertise 800 Gb/s connectivity.

Start a selection discussion with the traffic role and protocol. The Quantum-3 hardware guide describes 144 XDR data ports across 72 OSFP cages on Q3400-RA. The SN5610 ordering table describes 64 OSFP data cages, two SFP28 ports and Cumulus Linux authentication. Cage count, logical link count and supported endpoint rate must be kept separate when calculating topology and cable quantities.

Next check the facilities design. Q3400-RA is the AC-powered configuration; Q3401-RD is a different order code using a DC busbar. SN5610 has its own fan, rail and power-supply entries. Build the bill of materials from the chosen chassis rather than a family label.

This example establishes a documented reference use, not a customer deployment or a SwitchInfra delivery claim. NVIDIA says the representative components must be finalized for the actual design. A project still needs its workload, topology, cable validation, software entitlement and service scope checked before procurement.

Scope: Direct exact-OPN reference-design evidence for Q3400-RA and SN5610; educational context only for other products.

Public references

• Major Components — NVIDIA SuperPOD with DGX B300 Systems, NVIDIA Quantum-X800 InfiniBand switching and AC Power Reference Architecture
• NVIDIA Q32xx and Q34xx XDR 800Gb/s InfiniBand Switch Systems User Manual | NVIDIA Q32xx and Q34xx XDR 800Gb/s InfiniBand Switch Systems User Manual
• Ordering Information | NVIDIA Spectrum-4 SN5000 2U Switch Systems Hardware User Manual
• Accessory and Replacement Parts | NVIDIA Spectrum-4 SN5000 2U Switch Systems Hardware User Manual