(858) 613-1818Quote list 0
Allocation required

8 GPU HGX B300 AI server, air cooled

Eight NVIDIA B300 GPUs with 288 GB HBM3e each on an HGX baseboard, NVLink and NVSwitch, 8 x 800 Gb/s ConnectX-8, in a 10U air cooled chassis.

GPU: NVIDIA HGX B300

Illustration of a 10U HGX GPU server with the lid open
GPUs
8 x NVIDIA B300, 288 GB HBM3e each, NVLink and NVSwitch
CPU
2 x AMD EPYC 9005 or 2 x Intel Xeon 6
Memory
Up to 3 TB DDR5 RDIMM across 24 or 32 slots
Storage
2 x M.2 boot, up to 8 x U.2 or E3.S NVMe
Networking
8 x 800 Gb/s ConnectX-8 (east west), 2 x 400 Gb/s BlueField-3 (north south)
Power
Redundant PSUs, about 14 kW per node planning figure

Overview

This is the standard building block for Blackwell Ultra training and inference clusters in air cooled data centers. The HGX B300 baseboard carries eight B300 GPUs with 288 GB of HBM3e each, joined by NVLink and NVSwitch so the node behaves as one 2.1 TB accelerator. ConnectX-8 networking is integrated on the baseboard, giving each GPU its own 800 Gb/s port into the cluster fabric.

We configure the host side around your workload: dual AMD EPYC 9005 or Intel Xeon 6 processors, DDR5 memory up to 3 TB, mirrored boot drives and up to eight NVMe data drives for local checkpoints and datasets. Frontend connectivity uses BlueField-3 DPUs or ConnectX-7 adapters at 400 Gb/s for storage and north south traffic.

Every node is tested, NVLink and fabric validated, and shipped with an acceptance report. Plan for about 14 kW per node and ten rack units, which usually means three or four nodes per rack in a conventional hall. If your facility cannot support that density, ask about the liquid cooled variant, or we can lay the nodes out at a lower density that your hall or colocation facility can take.

Use cases

  • LLM pretraining and continued pretraining
  • Reasoning model inference at scale
  • Multi node training clusters
  • Fine tuning of 70B to 400B parameter models

Specifications

Compute
GPUs8 x NVIDIA B300 on the HGX B300 baseboard
GPU memory288 GB HBM3e per GPU, 2.1 TB per node as published by NVIDIA
GPU interconnectFifth generation NVLink, 1.8 TB/s per GPU, NVSwitch
FP4 performance144 PFLOPS sparse, 108 PFLOPS dense per node
CPUs2 x AMD EPYC 9005 series or 2 x Intel Xeon 6, up to 128 cores per socket class
System memory24 or 32 x DDR5 RDIMM, up to 3 TB
Storage
Boot2 x M.2 NVMe, mirrored
DataUp to 8 x U.2 or E3.S NVMe, 7.68 TB to 30.72 TB each
RAIDSoftware RAID or optional hardware RAID for boot volume
Networking
GPU fabric8 x 800 Gb/s InfiniBand or RoCE, ConnectX-8, OSFP
Frontend2 x 400 Gb/s, BlueField-3 or ConnectX-7, QSFP112
ManagementDedicated 1 GbE BMC port, TPM 2.0
Physical
Form factor10U rackmount, air cooled, front to back airflow
Power suppliesRedundant 3,000 W class PSUs, 200 to 240 V input
Planning powerAbout 14 kW per node
Inlet temperatureUp to 35 C class, check site conditions

Figures describe the system class. Exact configuration, weights and power draw are confirmed on the quotation.

What's included

  • Rails, power cords and cable management arm
  • Burn in with multi hour GPU stress and NVLink validation, acceptance report
  • NVIDIA end user registration handled with the order
  • Operating system imaging, driver and fabric manager baseline on request
  • 3 year warranty, options to 5 years with advance replacement

Compliance

Compare it with the next class up or down

The nearest classes we quote are the 8 GPU HGX B300 AI server, liquid cooled and the 8 GPU HGX B200 AI server. If neither is quite right, tell us the workload and we specify the system between them.

Talk to an engineerRequest a quotation