NVIDIA DGX SuperPOD: Next Generation Scalable Infrastructure for AI Leadership - Reference Architecture featuring NVIDIA DGX GB200

Overview

The architecture is built around the Scalable Unit (SU) of eight DGX GB200 NVL72 racks. Each SU requires 1.2 MW of TDP, and a SuperPOD can scale to 128 racks (9,216 GPUs) and beyond, with the fabric design supporting up to 16 SUs. The document walks through the rack internals (compute trays, NVLink switch trays, power shelves), the InfiniBand compute fabric, the Ethernet storage and management fabrics, storage requirements and the recommended rack layout.

Each rack holds 18 compute trays and 9 NVLink switch trays that form a single 72-GPU NVLink domain. On the power side every rack carries eight power shelves, each with six 5.5 kW PSUs delivering up to 33 kW. Cooling is hybrid: GPUs and CPUs are direct-liquid-cooled while other components stay air-cooled, which means the liquid and air fractions must be sized separately for the CDU and the air side.

The one hard facility requirement is availability: the data centre should meet or exceed Uptime Tier 3, TIA-942-B Rated 3 or EN 50600 Availability Class 3, meaning concurrent maintainability and no single point of failure. This is a vendor reference. It puts its weight on network and software detail and leaves floor plans, redundant power topology and mechanical piping to the facility designer.

Key points

1. Each SU (8 DGX GB200 NVL72 racks) requires 1.2 MW TDP.
2. Facility should meet or exceed Uptime Tier 3, TIA-942-B Rated 3 or EN 50600 Class 3, with concurrent maintainability and no single point of failure.
3. Hybrid cooling: GPUs and CPUs direct-liquid-cooled, other components air-cooled.
4. Eight power shelves per rack, each with 6 x 5.5 kW PSUs delivering up to 33 kW.
5. 18 compute trays and 9 NVLink switch trays per rack form one 72-GPU NVLink domain.
6. Scales to 128 racks (9,216 GPUs) and beyond; fabric design supports up to 16 SUs.

Using it in design

• Define the cluster template with SU = 8 NVL72 racks at 1.2 MW IT; auto-place SUs in rows following the reference rack layout, allocating ~150 kW per compute rack plus network racks.
• When a project contains DGX SuperPOD, enforce Tier III / Rated-3 / Class 3 topology checks (concurrent maintainability, no single point of failure).
• Add the NVL72 rack with 8 x 33 kW power shelves and its liquid/air cooling split to the catalog, and base CDU and air-side sizing on that split.

Who should read it

Facility and power designers planning a hall for a DGX GB200 SuperPOD.

Related standards

• TIA-942-B
• EN 50600
• Uptime Institute Tier Standard: Topology

Publication details

Authors

NVIDIA

Publisher

NVIDIA

Year

2025

Document number

RA-11338-001 V01

Document language

English

Source type

Manufacturer

Access

Free to read

Licence

Copyright retained; citation only — the publisher’s terms apply to the original document

Citation