NVIDIA HGX B200
We build and run clusters of HGX B200, the eight-GPU Blackwell system before B300, from new systems or ones you already own.
- GPUs
- 8 × B200
- GPU memory
- 180 GB HBM3e each
- Network
- 400 Gb/s per GPU
- Power
- Up to 14.3 kW per system
- Cooling
- Air, front to back
NVIDIA HGX B200 specifications.
One system on NVIDIA's figures, with notes where it differs from B300.
- Generation
- BlackwellThe generation before Blackwell Ultra, which HGX B300 uses.
- GPUs
- 8 GPUs per nodeAll eight share one NVLink domain.
- GPU memory
- 180 GB HBM3e per GPU, 1.44 TB per system108 GB less per GPU than HGX B300. It sets how large a model one system can hold.
- NVLink
- 14.4 TB/s across 8 GPUs1.8 TB/s per GPU, unchanged in HGX B300.
- Scale-out network
- 400 Gb/s per GPU (ConnectX-7)Against 800 Gb/s per GPU on HGX B300. Less bandwidth for jobs that span systems, and a lighter fabric to build.
- Power
- Up to 14.3 kW per systemNVIDIA gives it as a ceiling, not a typical draw.
- Cooling
- AirNo facility water at the rack. Rooms built for air can take it.
- Air-cooled rack
- Set in the designThe count per rack follows what each position can feed and cool.
- Reference design
- 32 nodesThe scalable unit in NVIDIA's reference architecture for B200.
- Lead time
- Set in the proposalFor systems you already hold, the schedule depends on the site and the network, not on supply.
Figures are NVIDIA's: its specifications for the eight-GPU B200 system and its B200 reference architecture.
Fitting B200 into an air-cooled hall.
An example rack of four systems, keyed to the notes. The design sets the real count per rack.
- 1
Feeds sized to NVIDIA's maximum
We size each position to NVIDIA's figure, up to 14.3 kW per system, rather than to typical draw. Every system takes power from an A and a B strip on separate circuits.
- 2
Room air, no facility water
The fan wall draws cold-aisle air in at the front and exhausts it at the back. That's why B200 fits rooms built for air, provided the hot aisle is contained.
- 3
ConnectX-7 links to the leaf switches
One per GPU at 400 Gb/s, so eight fabric links leave each system. On systems you already own, we design the fabric around the NICs they came with.
- 4
Four systems, up to 57.2 kW
The drawing's count times NVIDIA's per-system maximum, as an example. If your positions can't feed that, we put fewer systems in each rack.
HGX B200 or HGX B300?
Two eight-GPU systems on air, at about the same power. B300 adds memory and scale-out bandwidth per GPU; B200 makes sense where the models fit and the systems are already bought.
| HGX B300 | HGX B200 | |
|---|---|---|
| Generation | Blackwell Ultra | Blackwell |
| GPU memory | 288 GB HBM3e per GPU | 180 GB HBM3e per GPU |
| NVLink | 1.8 TB/s per GPU | 1.8 TB/s per GPU |
| Scale-out network | 800 Gb/s per GPU (ConnectX-8) | 400 Gb/s per GPU (ConnectX-7) |
| Power | About 14 kW per system | Up to 14.3 kW per system |
| Lead time | About 10 to 14 weeks from signing | Set in the proposal |
| Fits best | New builds, larger models and long context | Models that fit in 1.44 TB per system, and systems you already own |
NVIDIA's figures for both. If you already hold B200 systems, this table is about what to add next.
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 GB200 NVL72Blackwell72 GPUs per rack186 GB HBM3e per GPULiquidAvailable
- NVIDIA GB300 NVL72Blackwell Ultra72 GPUs per rack288 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
HGX B200 questions.
We build HGX B200 clusters as a deployment your team runs, a managed cluster you own and we operate, or a reserved cluster we own. Other platforms.
It depends on how many share the rack. NVIDIA gives up to 14.3 kW per system, so a rack of four draws up to 57.2 kW. We set the count from what your positions can feed.
B300 has 288 GB per GPU against 180, and 800 Gb/s of network per GPU against 400. NVLink per GPU and power per system are about the same.
If your models fit in a system's 1.44 TB of GPU memory and your jobs don't lean on traffic between systems, yes. For a new build you plan to run for years, compare it with HGX B300 first.
Yes. We inventory them, check firmware and health, add the network they're missing and burn the cluster in. Then your team takes it over, or we run it as a managed cluster you still own.
No. It cools with room air, front to back, so a hall without facility water can take it. The hot aisle still needs to be contained.
ConnectX-7, one per GPU at 400 Gb/s. We build it into a rail-optimized spine-leaf RoCE v2 Ethernet fabric, with storage and management on networks of their own.
64, in 8 nodes. NVIDIA's B200 reference architecture scales in units of 32 nodes.
It depends on the site and what has to be added: power, cooling, the network. The proposal sets the schedule once we've inventoried the systems and seen the hall.