(858) 613-1818Quote list 0

NVIDIA HGX B300

NVIDIA's Blackwell Ultra 8 GPU baseboard with 288 GB HBM3e per GPU, 1.8 TB/s NVLink per GPU and integrated 800 Gb/s ConnectX-8 networking.

Family
Blackwell Ultra
Edition
HGX 8 GPU baseboard
Memory
288 GB HBM3e per GPU, 2.1 TB total across the 8 GPU baseboard as published by NVIDIA
Memory bandwidth
8 TB/s per GPU
Interconnect
Fifth generation NVLink, 1.8 TB/s per GPU, 14.4 TB/s total NVLink bandwidth through NVSwitch across 8 GPUs
Form factor
SXM module on the HGX B300 baseboard with 8 integrated ConnectX-8 SuperNICs
TDP
Up to 1,400 W per GPU, as reported in NVIDIA's Blackwell Ultra technical material; not on the HGX product page. Air cooled systems run at a lower configured cap
Cooling
Air or liquid
Best for
Frontier LLM training, reasoning model inference and large multi node clusters
Availability
Allocation required

HGX B300 is the Blackwell Ultra generation of NVIDIA's eight GPU baseboard. Each GPU carries 288 GB of HBM3e at 8 TB/s, and the eight GPUs are joined by fifth generation NVLink and NVSwitch at 1.8 TB/s per GPU, so a single node holds a trillion parameter model in memory and serves it as one accelerator. NVIDIA publishes 144 PFLOPS of sparse FP4 and 72 PFLOPS of FP8 per baseboard, with dense FP4 1.5 times higher than HGX B200.

The baseboard integrates eight ConnectX-8 SuperNICs at up to 800 Gb/s each, so the east west fabric is part of the platform rather than an add in card. We supply HGX B300 in air cooled 10U chassis for standard racks at about 14 kW per node, and in liquid cooled chassis for high density halls. Both are quoted with AMD EPYC 9005 or Intel Xeon 6 processors, DDR5 memory, NVMe storage and BlueField-3 frontend networking.

Systems that carry this GPU

Related guides

Air cooled vs liquid cooled AI servers

How to decide between air and direct liquid cooling for GPU servers, from the rack power limit up to NVL72 racks that only come liquid cooled.

5 min read

In this guide

  1. The question is the rack, not the server
  2. How air cooling works at high density

Buying data center GPUs, export and registration

The compliance steps that come with every HGX, H200 NVL, Server Edition and NVL72 purchase, explained so your order ships on time.

4 min read

In this guide

  1. Why a GPU purchase comes with paperwork
  2. Which products are covered

Choosing a GPU server for LLM inference

A practical method for sizing an inference server from the model, the quantization and the concurrency you expect, with worked examples.

5 min read

In this guide

  1. Start with the model, not the server
  2. Step 1: how much GPU memory the weights need

HGX vs PCIe GPU servers

The real differences between an HGX node and a PCIe GPU server, and a simple rule for choosing between them.

5 min read

In this guide

  1. Two kinds of 8 GPU server
  2. The interconnect is the point

How to request a quotation

A short guide to getting a complete, accurate quotation, from the details we need to what the quotation contains.

4 min read

In this guide

  1. Why there are no prices on this site
  2. Three ways to ask

Sizing CPU, memory and storage around GPUs

Six rules of thumb that keep the GPUs fed, with the arithmetic for an 8 GPU node, a PCIe inference server and a 16 node training pod.

5 min read

In this guide

  1. The GPUs are the expensive part, so feed them
  2. The six rules

Want this GPU in a different chassis?

We quote every class that carries it, from a single node to a pod. Tell us the workload and we propose the chassis.

Request a quotationView quote list