NVIDIA GB300 NVL72
A liquid-cooled rack of 72 Blackwell Ultra GPUs and 36 Grace CPUs, which we fit to a site, build into clusters and operate.
- GPUs
- 72 GPUs, 36 Grace CPUs
- GPU memory
- 20.7 TB HBM3e per rack
- Network
- 800 Gb/s per GPU
- Power
- Up to 142 kW per rack
- Cooling
- Liquid, at the rack
NVIDIA GB300 NVL72 specifications.
The figures for one rack, with what each one asks of the hall.
- Generation
- Blackwell Ultra
- GPUs
- 72 GPUs per rackAll in one NVLink domain, so a single model can span the whole rack.
- CPUs
- 36 Grace per rackOne Grace for every two GPUs.
- GPU memory
- 288 GB HBM3e per GPU, 20.7 TB per rackA model that fits in this stays on NVLink, inside the rack.
- NVLink
- 130 TB/s across 72 GPUs1.8 TB/s per GPU, through switch trays inside the rack.
- Scale-out network
- 800 Gb/s per GPU (ConnectX-8)57.6 Tb/s per rack. It joins racks to each other, so it's sized to the number of racks you plan to reach.
- Power
- Up to 142 kW per rackNVIDIA's figure for one full rack. Network, storage and the cooling plant draw on top.
- Cooling
- Liquid, at the rackCold plates on the trays, through a coolant distribution unit on facility water. Some heat still goes to air.
- Loaded weight
- About 1.4 to 1.6 metric tonsVaries by build. We check it against the floor rating at each position.
- Reference designs
- 2, 4 and 8 racks144, 288 and 576 GPUs. At eight racks that's up to 1,136 kW before the network.
- Lead time
- Set in the proposalCounted from signing. It follows rack allocation and how ready the site is.
Figures are NVIDIA's, from its GB300 NVL72 product page and reference architecture. Loaded weight is from server manufacturers' documents.
What one rack asks of the room.
It arrives as a single system, so its position needs the power, the water and the floor for all of it at once. The notes follow the numbers on the drawing.
- 1
Up to 142 kW per rack
Fed on two independent paths, each sized for the rack's peak. Halls built for air often fall short of that at a single position.
- 2
Facility water at the position
A coolant distribution unit, in the rack or in the row, moves heat from the cold plates to the facility loop. Power shelves and NICs still shed some heat to the air.
- 3
800 Gb/s per GPU between racks
57.6 Tb/s of ConnectX-8 links leave each rack for the leaf switches. Inside the rack, NVLink carries the GPU traffic.
- 4
About 1.4 to 1.6 metric tons, loaded
That is more than many raised floors are rated for. The dock, the lift and each doorway on the way in get checked too.
GB300 NVL72 or GB200 NVL72?
The same rack shape, NVLink domain and floor load. The difference is in memory and network per GPU, and in what each position has to feed.
| GB200 NVL72 | GB300 NVL72 | |
|---|---|---|
| Generation | Blackwell | Blackwell Ultra |
| GPU memory | 186 GB HBM3e per GPU | 288 GB HBM3e per GPU |
| Memory per rack | 13.4 TB per rack | 20.7 TB per rack |
| NVLink | 130 TB/s across 72 GPUs | 130 TB/s across 72 GPUs |
| Scale-out network | 400 Gb/s per GPU (ConnectX-7) | 800 Gb/s per GPU (ConnectX-8) |
| Power | About 120 kW per rack | Up to 142 kW per rack |
| Loaded weight | About 1.4 to 1.6 metric tons | About 1.4 to 1.6 metric tons |
| Reference designs | 8 racks | 2, 4 and 8 racks |
| Fits best | Workloads that fit its memory, on sites already fed for it | Larger models and longer context, on sites that can feed the extra draw |
Figures from NVIDIA, weights from manufacturers. A position fed for GB200 may need new feeds for GB300.
Other NVIDIA platforms we build on.
The same four stages on each: design, source, build and operate. All hardware.
- NVIDIA HGX B300Blackwell Ultra8 GPUs per node288 GB HBM3e per GPUAirAvailable
- NVIDIA HGX B200Blackwell8 GPUs per node180 GB HBM3e per GPUAirAvailable
- NVIDIA GB200 NVL72Blackwell72 GPUs per rack186 GB HBM3e per GPULiquidAvailable
- NVIDIA HGX Rubin NVL8Rubin8 GPUs per node288 GB HBM4 per GPULiquidWaitlist
- NVIDIA VR200 NVL72Rubin72 GPUs per rack288 GB HBM4 per GPULiquidWaitlist
GB300 NVL72 questions.
We build GB300 NVL72 clusters as a deployment your team runs, a managed cluster you own and we operate, or a reserved cluster we own. Other platforms.
NVIDIA gives up to 142 kW per rack. We feed each position for that peak on two independent paths, then add the network, storage and cooling plant.
Yes. The compute trays are cold-plate cooled and can't run on air alone. The hall needs facility water at the position, and air handling for the heat that's left.
About 1.4 to 1.6 metric tons loaded, depending on the build. Many raised floors aren't rated for that, so some sites set NVL72 racks directly on the slab.
Usually the integrated rack: compute trays, NVLink switch trays and power shelves, tested at the factory. The network between racks, coolant distribution, freight and installation are often quoted separately. We scope all of it in one proposal.
One rack (72 GPUs) is the smallest we build. NVIDIA validates designs at 2, 4 and 8 racks, or 144, 288 and 576 GPUs, and we scope larger ones case by case.
Yes, where the hall can bring facility water to the position and feed the rack's power. Where it can't, we look at retrofitting the hall or building a modular site.
It is when your models need 288 GB per GPU or 800 Gb/s per GPU between racks. If the workload fits GB200's memory and the site can't feed the higher draw, GB200 NVL72 is the easier fit.
The proposal sets the dates, counted from signing. They follow rack allocation and what the site needs first: power, facility water or floor work.