GS-SYS-08Engineering datasheet

NVIDIA HGX B300

Statusshipping
Verified2026-07-15
Download datasheet (PDF) ↓

HGX B300 is the 8-GPU Blackwell Ultra baseboard NVIDIA sells to OEMs, the platform under every non-DGX B300 server. It carries 2.3 TB of HBM3e across eight 288 GB GPUs on a 14.4 TB/s NVLink 5 fabric, and unlike HGX B200 it integrates the scale-out network on the platform itself: eight ConnectX-8 SuperNICs at 800 Gb/s per GPU over PCIe Gen6, so the NIC choice OEMs used to make is now fixed by NVIDIA. Supermicro, Dell, HPE, Lenovo, Aivres and others wrap their own CPUs, chassis and cooling around it; the same baseboard ships in air-cooled 8U boxes at the 1,100 W GPU bin and in 2-OU/4U direct-liquid chassis, with the full 1,400 W bin reserved for liquid builds. It is the default OEM quote for new Blackwell capacity in 2026.

8x B300 SXM
GPUs · Blackwell Ultra, fully NVLinked
2.3 TB HBM3e total
GPU memory · 288 GB per GPU
up to 64 TB/s aggregate
Memory bandwidth · 8 TB/s per GPU
14.4 TB/s aggregate
NVLink · NVLink 5, 1.8 TB/s per GPU
8x ConnectX-8 on-platform
Networking · 800 Gb/s per GPU, PCIe Gen6
1,100 W air / up to 1,400 W liquid
GPU TDP · per GPU, bin set by OEM chassis
Compute01/05
GPUNVIDIA B300 (Blackwell Ultra) · SXM form factor, 8 per baseboard
FP4 Tensorup to 144 PFLOPS · 8-GPU platform, NVIDIA figure; dense is 120 PFLOPS
FP8 Tensor72 PFLOPS · 8-GPU platform, NVIDIA figure
Attention throughput~2x HGX B200 · doubled SFU throughput
GPU TDP1,100 W (air) / up to 1,400 W (liquid) · configurable per GPU
Memory02/05
Memory typeHBM3e, 12-Hi stacks · per GPU
Memory per GPU288 GB · 50% more than HGX B200's 180 GB
Total baseboard memory2.3 TB · 8x 288 GB
Bandwidth per GPUup to 8 TB/s
Aggregate bandwidthup to 64 TB/s · per node
Interconnect03/05
NVLink generationNVLink 5 · same generation as HGX B200
GPU-to-GPU bandwidth1.8 TB/s per GPU · bidirectional
Aggregate NVLink bandwidth14.4 TB/s · across all 8 GPUs
Switch fabricNVSwitch · full any-to-any GPU mesh
Scale-out NICs8x ConnectX-8 SuperNICs, on-platform · 800 Gb/s per GPU (2x 400G), PCIe Gen6 direct to GPU
vs. HGX B200NICs integrated, not OEM-selected · B200 platforms left NIC choice to the OEM
Power & cooling04/05
GPU-only draw (air)8.8 kW · 8x 1,100 W bin
GPU-only draw (liquid)up to 11.2 kW · 8x 1,400 W bin
Full server draw12 to 16 kW · adds CPUs, DRAM, NVMe, NICs, fans; DGX B300 reference is 14.5 kW
Cooling optionsAir (8U class) or direct-to-chip liquid (2-OU/4U class) · liquid unlocks the top TDP bin and rack density
OEM platforms05/05
Ships fromSupermicro, Dell, HPE, Lenovo, Aivres, and other NVIDIA-certified OEMs · each builds its own chassis around the baseboard
Example SKUsSupermicro SYS-822GS-NBR (8U air), SYS-422GS-NB3RT-ALC (4U liquid), SYS-222GS-NB3OT-ALC (2-OU liquid)
CPU choiceAMD EPYC 9005/9004 or Intel Xeon · OEM- and SKU-dependent
Baseboard part number935-26287-0070-000 · sold as a discrete part through distribution
vs. DGX B300Same platform, OEM picks chassis/CPU/cooling · DGX is NVIDIA's fixed-config, air-cooled-only equivalent
Field notes
  • HGX B300 is a baseboard, not a server, but it fixes more of the design than HGX B200 did: GPUs, NVSwitch fabric and the eight ConnectX-8 SuperNICs are all NVIDIA's. The OEM now differentiates only on CPU, memory, storage, chassis and cooling.
  • Air-cooled B300 exists (1,100 W bin, 8U class) despite widespread claims that Blackwell Ultra is liquid-only. What is true: the full 1,400 W bin needs direct liquid, and four 8U air boxes per rack already demand ~60 kW rack power, so air works only in high-density halls.
  • The 800 Gb/s per-GPU scale-out only pays off on an 800G-class fabric (Quantum-X800 or Spectrum-X SN5600-class). Budget switches and optics with the node order; at 8x 800 Gb/s per node the fabric BOM rivals mid-five-figures per node.
  • Per-GPU HBM capacity, not FLOPS, is the usual reason to specify B300 over B200: 288 GB holds model shards and KV cache that would force tensor-parallel spill on 180 GB parts, and that single change often removes a whole parallelism dimension from the deployment.
  • Two 'HGX B300' servers from different OEMs still differ meaningfully in CPU platform, DRAM ceiling, storage layout and serviceability. Compare full system BOMs, not the shared baseboard spec sheet.
Questions we get on this part

How much does an HGX B300 server cost?

Reported street pricing for a fully configured 8-GPU HGX B300 server runs roughly $430,000 to $550,000 or more, with Aivres quoting from about $430,000. This excludes the 800G switch fabric, optics, and rack power and cooling infrastructure.

What is the difference between HGX B300 and HGX B200?

B300 carries 288 GB per GPU (2.3 TB per baseboard) versus B200's 180 GB (1.44 TB), raises the GPU power ceiling to 1,400 W liquid-cooled, adds roughly 1.5x the dense FP4 throughput, and integrates eight ConnectX-8 SuperNICs at 800 Gb/s per GPU on the platform, where HGX B200 left NIC selection to the OEM.

Is there an air-cooled HGX B300 server?

Yes. Supermicro ships an 8U air-cooled HGX B300 system (SYS-822GS-NBR) running the GPUs at the 1,100 W TDP bin, alongside 2-OU and 4U direct-liquid systems. The full 1,400 W GPU bin is only available in liquid-cooled chassis.

What is the difference between HGX B300 and DGX B300?

Same 8-GPU Blackwell Ultra platform. HGX B300 is the baseboard OEMs build their own servers around, with choice of CPU, chassis and air or liquid cooling. DGX B300 is NVIDIA's fixed-configuration, air-cooled 10U appliance on the same platform with NVIDIA-direct support.

What is the HGX B300 baseboard part number?

The 8-GPU HGX B300 baseboard circulates in distribution as NVIDIA part number 935-26287-0070-000, though in practice nearly all volume ships inside complete OEM servers rather than as bare baseboards.