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

Hardware · Next generation

NVIDIA VR200 NVL72

NVIDIA's Vera Rubin NVL72 rack is on a waitlist for first allocation, and we design the site before the racks arrive.

Sheet 01 · VR200 NVL7272 GPUs per rack
GPUs
72 × Rubin
GPU memory
20.7 TB HBM4 per rack
NVLink
260 TB/s across 72 GPUs
Rack power
Not published
Status
Waitlist
Drawing is illustrative · Figures are NVIDIA's
Specifications

NVIDIA VR200 NVL72 specifications.

The published figures, and the ones a site plan still waits on.

Generation
Rubin
GPUs
72 GPUs per rackEvery GPU in the rack shares one NVLink domain. The scale-out network only joins racks to each other.
CPUs
36 Vera per rackVera in place of Grace, one CPU for every two GPUs.
GPU memory
288 GB HBM4 per GPU, 20.7 TB per rackThe same capacity as GB300 NVL72. The step is from HBM3e to HBM4.
NVLink
260 TB/s across 72 GPUs3.6 TB/s per GPU, against 1.8 TB/s per GPU on GB300 NVL72.
Scale-out network
Confirmed with the orderWe confirm the NIC with the order. GB300 NVL72 runs 800 Gb/s per GPU, and we plan cabling for at least that.
Cooling
Liquid, at the rackAs on GB300 NVL72, every position needs facility water. The supply temperature is set in the design.
Rack power
Not publishedGB300 NVL72 draws up to 142 kW per rack. Until NVIDIA gives a figure, we plan above that.
Rack weight
Confirmed with the orderA loaded GB300 NVL72 rack is about 1.4 to 1.6 metric tons. The floor is planned for at least that.
Availability
Waitlist for first allocationThere's no lead time yet; the proposal sets one.

VR200 NVL72 figures are from NVIDIA's Vera Rubin NVL72 page. Where a figure isn't set yet, the note gives the GB300 NVL72 figure we plan from.

Site requirements

What to have in the hall before allocation.

One liquid-cooled rack in front elevation. NVIDIA hasn't published a rack power figure for VR200 NVL72, and network and weight are confirmed with the order. Each note says what we plan around until then.

  1. 1

    Power sized above GB300 NVL72

    GB300 NVL72 draws up to 142 kW per rack. With no published figure for VR200 NVL72, we size feeds, busway and transformers above that and check the design against the order.

  2. 2

    Facility water at every position

    Like GB300 NVL72, the rack is cooled by liquid, so each position needs facility water and coolant distribution. We set the supply temperature in the design, against NVIDIA's figures for the order.

  3. 3

    Cable paths with room to spare

    The NIC and the fabric are confirmed with the order. Until then, cable paths and leaf positions are sized for at least 800 Gb/s per GPU, the GB300 NVL72 figure.

  4. 4

    A floor rated for a full rack

    A GB300 NVL72 rack weighs about 1.4 to 1.6 metric tons loaded, so we rate the floor for at least that and confirm it against the shipping weight.

Choosing a platform

Vera Rubin NVL72 next to GB300 NVL72.

Both are 72-GPU liquid-cooled racks, a generation apart. VR200 NVL72 brings Vera CPUs, HBM4 and twice the NVLink bandwidth; GB300 NVL72 can be built now.

Vera Rubin NVL72 next to GB300 NVL72.
GB300 NVL72VR200 NVL72
GenerationBlackwell UltraRubin
CPUs36 Grace per rack36 Vera per rack
GPU memory288 GB HBM3e per GPU288 GB HBM4 per GPU
NVLink130 TB/s across 72 GPUs260 TB/s across 72 GPUs
Scale-out network800 Gb/s per GPU (ConnectX-8)Confirmed with the order
Rack powerUp to 142 kW per rackNot published
Rack weightAbout 1.4 to 1.6 metric tonsConfirmed with the order
AvailabilityDates set in the proposalWaitlist for first allocation

Both are quoted per project. A hall built with headroom can take GB300 NVL72 racks now and VR200 NVL72 racks once allocation arrives.

Other platforms

Other NVIDIA platforms we build on.

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

Questions

VR200 NVL72 questions.

We build VR200 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