Vertiv High Density Reference Design for AI #0019: 1.2 MW NVIDIA GB200 NVL72 SuperPOD (N redundancy)

Overview

The design carries 1,176 kW (about 1.2 MW) across 18 racks: eight 132 kW compute racks and ten 12 kW support racks. Cooling is 76% direct-to-chip liquid and 24% from a chilled-water modular fan wall. Infrastructure redundancy is deliberately N, meaning single points of failure are accepted. That makes the document useful as a picture of the minimum system you get once the cost of redundancy is stripped out.

The power chain is lean: a single 1,500 kVA UPS carries the IT load behind a 2,500 A switchboard. A separate 250 kVA UPS serves the mechanical load, running N+1, with a second unit added if the 2N option is chosen. Distribution uses 2 x 500 A busbar per row and 2 x 120 A tap-off boxes per rack, while network and storage racks are fed from 22 kW (32 A) PDUs.

One physical detail feeds straight into layout: the racks are provisioned for a 200 mm depth extension. Manifolds, hoses and cabling in a liquid-cooled rack do not fit in standard rack depth, and the document makes that allowance a design assumption from the start. The prefabricated-module approach also aims to cut on-site labour and commissioning time. The limit is plain: N redundancy only suits environments that do not need maintenance or fault tolerance, or that provide resilience at the software layer. A Tier-style availability target would call for the 2N or 4-to-make-3 templates. The document is copyright vendor material, so its numbers can be cited but its text not copied.

Key points

1. IT load 1.2 MW (1,176 kW): 8 x 132 kW + 10 x 12 kW, 18 racks
2. Cooling 76% liquid, 24% air via chilled-water fan wall
3. Infrastructure redundancy N; IT UPS 1 x 1,500 kVA; 2,500 A switchboard
4. Mechanical UPS 250 kVA (second unit for 2N)
5. Busbar 2 x 500 A per row; 2 x 120 A tap-off per rack; 22 kW (32 A) PDUs for network/storage racks
6. Racks provisioned for 200 mm depth extension

Using it in design

• In the layout engine add 200 mm to rack depth for liquid-cooled AI racks; manifolds and hoses do not fit standard depth.
• Offer an "N (non-redundant)" AI template next to 2N and 4-to-make-3 templates to show the cost versus availability trade-off.
• For a 1.2 MW block, a first sizing check pairs a 1,500 kVA IT UPS, a 250 kVA mechanical UPS and a 2,500 A switchboard.
• Define 2 x 500 A busbar with 2 x 120 A tap-offs per rack as the default AI row configuration.

Who should read it

Designers and budget owners building minimum-cost AI infrastructure at single-SuperPOD scale.

Related standards

• OCP Open Rack V3
• ASHRAE TC9.9 Liquid Cooling Guidelines

Publication details

Authors

Vertiv

Publisher

Vertiv

Year

2024

Document number

AI Reference Design #0019

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