Aircloud Compute is now Lightyard. Aircloud continues as a private inference platformAircloud continues for inference

Hardware · NVL72

NVIDIA GB200 NVL72

GB200 NVL72 puts 72 Blackwell GPUs in one liquid-cooled rack. We build new clusters of them and take over racks that are already installed.

Sheet 01 · GB200 NVL7272 GPUs per rack
GPUs
72 GPUs, 36 Grace CPUs
GPU memory
13.4 TB HBM3e per rack
NVLink
130 TB/s across 72 GPUs
Power
About 120 kW per rack
Cooling
Liquid, at the rack
Drawing is illustrative · Figures are NVIDIA's
Specifications

NVIDIA GB200 NVL72 specifications.

The rack as NVIDIA publishes it, and what each figure means on site.

Generation
BlackwellGB300 NVL72 keeps the same rack layout on Blackwell Ultra.
GPUs
72 GPUs per rackOne NVLink domain, so the rack arrives, installs and runs as one system.
CPUs
36 Grace per rackEach Grace CPU is paired with two GPUs.
GPU memory
186 GB HBM3e per GPU, 13.4 TB per rackA model split across all 72 GPUs has the full 13.4 TB to work with over NVLink.
NVLink
130 TB/s across 72 GPUs1.8 TB/s per GPU. Traffic inside the rack stays off the scale-out fabric.
Scale-out network
400 Gb/s per GPU (ConnectX-7)A rack has 72 of them, 28.8 Tb/s in total.
Power
About 120 kW per rackThe rack alone. The fabric and storage add their own load.
Cooling
Liquid, at the rackLiquid takes most of the heat. The rest leaves as air.
Loaded weight
About 1.4 to 1.6 metric tonsFrom server manufacturers' documents. It varies with the build.
Reference design
8 racksNVIDIA's scalable unit, 576 GPUs.
Lead time
Set in the proposalRacks already installed skip this. A takeover starts with an assessment.

Figures are NVIDIA's for the GB200 NVL72 rack, except loaded weight, which comes from server manufacturers' documents.

Site requirements

What a rack position needs before power-on.

One rack in elevation, numbered to the notes. On racks already installed, these are the four points a takeover assessment starts from.

  1. 1

    About 120 kW per rack

    The rack's power shelves take two independent feeds. We size both for peak draw and prove them under load during burn-in.

  2. 2

    A coolant loop commissioned first

    A coolant distribution unit, in the rack or in the row, feeds cold plates on every compute tray. The loop is flushed, filled and held at test pressure, and every quick disconnect is checked, before a tray powers on.

  3. 3

    72 ConnectX-7 links to the leaf switches

    One per GPU at 400 Gb/s, on a spine-leaf RoCE v2 Ethernet fabric. They carry the traffic between racks; NVLink handles everything within one.

  4. 4

    About 1.4 to 1.6 metric tons loaded

    All of it on one rack footprint. We check the floor rating at each position, and the route in from the loading dock, before the racks ship.

Choosing a platform

GB200 NVL72 against GB300 NVL72.

Both put 72 GPUs in one NVLink domain per rack. GB300 NVL72 adds memory and network per GPU, at more power per rack.

GB200 NVL72 against GB300 NVL72.
GB300 NVL72GB200 NVL72
GenerationBlackwell UltraBlackwell
GPU memory288 GB HBM3e per GPU186 GB HBM3e per GPU
Memory per rack20.7 TB per rack13.4 TB per rack
Scale-out network800 Gb/s per GPU400 Gb/s per GPU
PowerUp to 142 kW per rackAbout 120 kW per rack
Reference designs2, 4 and 8 racks8 racks
Fits bestNew builds, and models that need the most memory per GPUModels that fit its memory, and installed racks that need an operator

Figures from NVIDIA. Either is quoted per project, built new or taken over.

Other platforms

Other NVIDIA platforms we build on.

The same four stages on each: design, source, build and operate. All hardware.

Questions

GB200 NVL72 questions.

We build GB200 NVL72 clusters as a deployment your team runs, a managed cluster you own and we operate, or a reserved cluster we own. Other platforms.

Get started

Tell us where you're starting.

Capital, hardware, a site or a team: tell us what you have and when you need it running. We usually reply within a few hours and come back with a written plan.

sales@lightyard.ai