NVIDIA-DGX-SuperPOD-GB200-RA

NVIDIA DGX SuperPOD GB200 Reference Architecture

Type: Reference Architecture Tags: NVIDIA, DGX SuperPOD, DGX GB200, GB200 NVL72, Blackwell, Grace Blackwell, NVLink 5, NDR InfiniBand, Spectrum-4, Mission Control, AI factory Related: NVIDIA-DGX-SuperPOD, NVIDIA-GB200-NVL72, NVIDIA-DGX, NVIDIA-DGX-SuperPOD-B200-RA, NVIDIA-DGX-B200, NVIDIA-DGX-B300, NVIDIA-GB300-NVL72, NVIDIA-Blackwell-Architecture, NVIDIA-Grace-CPU, NVLink, NVIDIA-Quantum-InfiniBand, NVIDIA-Spectrum-X, NVIDIA-Mission-Control, NVIDIA-AI-Enterprise, NGC, NVIDIA-UFM, NVIDIA-Certified-Storage, NVIDIA-Enterprise-AI-Factory Sources: https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-gb200/latest/index.html, https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-gb200/latest/abstract.html, https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-gb200/latest/dgx-superpod-components.html, https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-gb200/latest/network-fabrics.html, https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-gb200/latest/dgx-software.html Last Updated: 2026-05-09

Summary

NVIDIA DGX SuperPOD GB200 Reference Architecture is the NVIDIA DGX SuperPOD design for rack-scale NVIDIA-GB200-NVL72 systems powered by Grace CPUs and Blackwell GPUs. It introduces a SuperPOD pattern with eight DGX GB200 rack systems per scalable unit, NVLink 5 inside each rack, NDR 400 Gb/s InfiniBand for cross-rack compute, Spectrum-4/Spectrum-X Ethernet for storage and in-band management, NVIDIA-Mission-Control, NVIDIA-AI-Enterprise, NGC, Magnum IO, Slurm, and NVIDIA Enterprise Support.

Detail

Purpose

Use this page for the NVIDIA document titled “NVIDIA DGX SuperPOD: Next Generation Scalable Infrastructure for AI Factories Reference Architecture Featuring NVIDIA DGX GB200.” Use NVIDIA-GB200-NVL72 for the rack-scale system identity and NVIDIA-DGX-SuperPOD for the generic SuperPOD family.

Architecture notes

• Document number RA-11338-001 V01; publication date 2025-06-16; docs last updated 2025-11-19.
• The RA is based on DGX GB200 systems powered by NVIDIA-Grace-CPU and Blackwell GPUs.
• Each scalable unit contains eight DGX GB200 systems.
• Each DGX GB200 rack system is a 72-GPU NVLink domain with 18 compute trays and nine NVLink switch trays.
• Each compute tray contains two GB200 Superchips, and each superchip combines two B200 GPUs and one Grace CPU over NVLink-C2C.
• The architecture includes five network fabrics: NVLink 5, compute fabric, storage fabric, in-band management network, and out-of-band management network.
• Physical fabrics include multi-node NVLink, compute InfiniBand, storage/in-band Ethernet, and out-of-band networking.
• Compute fabric uses NVIDIA Quantum QM9700 NDR 400 Gb/s InfiniBand in a rail-optimized full-fat-tree design.
• Storage and in-band management use NVIDIA Spectrum-4 SN5600 Ethernet switches with 800 Gb/s class Ethernet.
• Storage guidance uses RDMA over Converged Ethernet v2 for high-performance storage from DGX SuperPOD certified storage partners.

Software and operations

NVIDIA-Mission-Control is the management and operations stack for DGX GB200 SuperPOD deployments. The software components include Mission Control, NVIDIA-AI-Enterprise, Magnum IO, NGC, and Slurm. The RA also calls out building-management-system integration because of the power and cooling complexity of the rack-scale GB200 design.

NVIDIA context

This page captures the Blackwell rack-scale SuperPOD reference design. It connects NVIDIA-DGX-SuperPOD-B200-RA to Blackwell Ultra pages such as NVIDIA-GB300-NVL72 and NVIDIA-DGX-B300, and it is especially useful when a customer needs the design implications of rack-scale NVLink domains rather than 8-GPU DGX nodes.

Connections

• NVIDIA-DGX-SuperPOD - parent SuperPOD architecture family.
• NVIDIA-GB200-NVL72 - rack-scale Grace Blackwell system used by this RA.
• NVIDIA-DGX - broader DGX systems family.
• NVIDIA-DGX-SuperPOD-B200-RA - preceding B200 SuperPOD architecture.
• NVIDIA-DGX-B200 - node-level Blackwell DGX comparison point.
• NVIDIA-DGX-B300 and NVIDIA-GB300-NVL72 - Blackwell Ultra successor direction.
• NVIDIA-Blackwell-Architecture - GPU architecture foundation.
• NVIDIA-Grace-CPU - CPU component in Grace Blackwell superchips.
• NVLink - NVLink 5 creates the 72-GPU rack-scale domain.
• NVIDIA-Quantum-InfiniBand - NDR InfiniBand compute fabric.
• NVIDIA-Spectrum-X - Spectrum-4/Spectrum-X Ethernet storage and management fabric direction.
• NVIDIA-Mission-Control - operations and orchestration stack for DGX GB200 SuperPOD.
• NVIDIA-AI-Enterprise - enterprise AI software layer.
• NGC - GPU-optimized software catalog.
• NVIDIA-UFM - InfiniBand fabric management and telemetry.
• NVIDIA-Certified-Storage - storage partner validation is required for SuperPOD storage.
• NVIDIA-Enterprise-AI-Factory - the RA is a practical AI factory build pattern.

Source Excerpts

• NVIDIA describes the GB200 RA as a next-generation AI factory architecture based on DGX GB200 systems with Grace CPUs and Blackwell GPUs.
• The abstract says each scalable unit contains eight DGX GB200 systems.
• The key-components page lists DGX GB200 NVL72 racks, NVLink 5, Quantum QM9700 InfiniBand, Spectrum-4 SN5600 Ethernet, Mission Control, AI Enterprise, NGC, Magnum IO, and Slurm.

Resources

• NVIDIA DGX SuperPOD GB200 Reference Architecture