NVIDIA HGX B300
An eight-GPU Blackwell Ultra system that runs on air. We design, build and operate clusters of them.
- GPUs
- 8 × B300
- GPU memory
- 288 GB HBM3e each
- Network
- 800 Gb/s per GPU
- Power
- About 14 kW per system
- Lead time
- About 10 to 14 weeks from signing
NVIDIA HGX B300 specifications.
What each eight-GPU system brings, and what it asks for.
- Generation
- Blackwell Ultra
- GPUs
- 8 GPUs per nodeLinked by NVLink inside the system.
- GPU memory
- 288 GB HBM3e per GPU, 2.3 TB per system
- NVLink
- 14.4 TB/s across 8 GPUsInside the system. Between systems, the scale-out fabric carries the traffic.
- Scale-out network
- 800 Gb/s per GPU (ConnectX-8)One NIC per GPU, so the fabric is sized per GPU rather than per system.
- Power
- About 14 kW per systemNVIDIA's figure for its eight-GPU B300 system.
- Cooling
- AirLiquid-cooled versions exist for halls with facility water.
- Air-cooled rack
- 4 systems, about 56 kWNVIDIA's reference layout. Your hall may call for fewer per rack.
- Reference designs
- 32, 64 and 128 nodesPod sizes in NVIDIA's reference architecture for HGX B300.
- Lead time
- About 10 to 14 weeks from signing
HGX B300 figures are NVIDIA's published specifications for its eight-GPU system.
What a rack of them asks of the hall.
An air-cooled rack, front elevation. Each note matches a number on the drawing.
- 1
Two power feeds per rack
Each system draws about 14 kW per system, fed from A and B power strips on separate circuits, so one feed can fail without taking the rack down.
- 2
Air in at the front, out at the back
Cold aisle in through the fan wall, hot air out the back into a contained hot aisle. At this density the hall needs hot-aisle containment.
- 3
800 Gb/s per GPU to the leaf switches
8 ConnectX-8 links per system, one per GPU, run to leaf switches in a network rack nearby. Storage and management networks run alongside.
- 4
4 systems to a rack, about 56 kW
That's NVIDIA's air-cooled layout. Where a position can't feed it, the same systems spread across more racks.
HGX B300 or GB300 NVL72?
The same GPU generation in a different unit. The choice usually comes down to how large your models are and what the site can take.
| GB300 NVL72 | HGX B300 | |
|---|---|---|
| Unit | 72 GPUs per rack | 8 GPUs per node |
| NVLink domain | 72 GPUs in one rack | 8 GPUs in one system |
| Power | Up to 142 kW per rack | About 14 kW per system |
| Cooling | Liquid, at the rack | Air, or liquid where the hall has it |
| Smallest cluster | One rack (72 GPUs) | 8 nodes (64 GPUs) |
| Fits best | Models split across many GPUs, on sites built for liquid cooling | Models that fit in eight GPUs, and halls without facility water |
Both are quoted per project. We recommend one once we know the workload and the site.
Other NVIDIA platforms we build on.
The same four stages on each: design, source, build and operate. All hardware.
- NVIDIA HGX B200Blackwell8 GPUs per node180 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 B300 questions.
We build HGX B300 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 about 14 kW per system for its eight-GPU B300 system. A cluster also draws power for its network, storage and management racks.
NVIDIA's air-cooled layout puts 4 in a rack at about 56 kW. Many halls can't feed that per position, so the layout often uses fewer per rack.
The standard system is. Liquid-cooled versions exist for halls with facility water, and they allow more systems per rack.
The smallest we build is 8 nodes. Most clusters are 32 to 128 nodes; larger ones are scoped case by case. NVIDIA's reference architecture validates pods of 32, 64 and 128 nodes.
One ConnectX-8 per GPU at 800 Gb/s per GPU, on a rail-optimized spine-leaf RoCE v2 Ethernet fabric, plus separate storage and management networks. We design all three.
About 10 to 14 weeks from signing for the hardware. Larger clusters are handed over in monthly waves of 32 nodes.
HGX B300 if your models fit in eight GPUs or your hall runs on air. GB300 NVL72 if you split models across many GPUs and the site can take liquid-cooled racks.