Skip to content
EMARQUE AI

Built to order

EMARQUE AI HGX B300

Eight NVIDIA Blackwell Ultra GPUs on one HGX B300 board, with 2.1 TB of GPU memory between them. Built to order for training large models, or serving one to a whole organisation.
  • 8 × NVIDIA Blackwell Ultra GPUs on an HGX B300 board
  • 2.1 TB HBM3e, NVLink at 1.8 TB/s GPU to GPU
  • Dual AMD EPYC or Intel Xeon, chosen in your quote
  • Up to 4 TB registered ECC DDR5
  • 8 × NVIDIA ConnectX-8 SuperNIC, 800 Gb/s per port
  • 12 × 3,000 W redundant power supplies
Get a quote

The price depends on the processors, memory, storage and networking specified around the eight-GPU board.

Air-cooled HGX B300 server, front three-quarter view

Product overview

Front of the HGX B300 server
Front: fan modules over the drive bays
Rear of the HGX B300 server
Rear: twelve power supplies, network and PCIe bays
The eight-GPU HGX B300 board on its tray, from above
The eight-GPU HGX B300 board on its tray

Key features

HGX B300 server with the GPU tray drawn out, showing the eight GPU modules

Eight GPUs working as one

Eight Blackwell Ultra GPUs sit on a single HGX B300 board with 2.1 TB of HBM3e between them, linked by NVLink at 1.8 TB/s. A model too large for one GPU runs across all eight.
The CPU tray from above: two processor sockets and the memory banks

Two processors, up to 4 TB ECC

Two AMD EPYC or Intel Xeon processors, chosen in your quote, with up to 4 TB of registered ECC memory, so preparing and loading data keeps eight GPUs fed.

  • 800 Gb/s per port

    Eight NVIDIA ConnectX-8 SuperNICs. That is the networking NVIDIA specifies for this board, and what a second machine needs if the work outgrows one.

  • Redundant power

    Twelve 3,000 W Titanium supplies, with spares, so the server keeps running if one fails.

  • Built to order

    There is no fixed configuration. We specify the processors, memory, storage and networking with you, source the platform through NVIDIA's OEM partners and set it up on site, under a three-year system warranty.

Use cases

  • Training

    Training and fine-tuning large models

    Models that will not fit on one GPU, trained or fine-tuned across all eight, with the networking to add a second machine later.

  • Inference

    One large model, served to everyone

    A single large model held in 2.1 TB of GPU memory and served from one machine, with room for long context windows.

  • Platform

    Shared capacity for several teams

    GPU capacity divided between business units or faculties, with isolation and scheduling handled at the platform layer.

Specifications

The final parts list is confirmed in your quote.

Air-cooled HGX B300 server, front three-quarter view
GPU
8 × NVIDIA Blackwell Ultra GPUs on an NVIDIA HGX B300 board
GPU memory
2.1 TB HBM3e across the eight GPUs (NVIDIA figure)
AI compute
144 PFLOPS FP4 with sparsity, 108 PFLOPS dense (NVIDIA figure)
GPU interconnect
NVLink, 1.8 TB/s GPU to GPU
CPU
Dual AMD EPYC or Intel Xeon, chosen in your quote
Memory
Up to 4 TB registered ECC DDR5
Storage
Up to 16 TB Gen5 NVMe
Networking
8 × NVIDIA ConnectX-8 SuperNIC, 800 Gb/s per OSFP port
Power supplies
12 × 3,000 W, redundant, 80 PLUS Titanium (installed capacity, not the draw)
Form factor
8U, 19-inch rackmount
Software
Ubuntu LTS. NVIDIA AI Enterprise is licensed separately
Warranty
3-year EMARQUE system warranty

Questions about the AI HGX B300

  • What is an HGX B300 board?
    NVIDIA's eight-GPU board: eight Blackwell Ultra GPUs mounted together and linked by NVLink, so they can work on one model at the same time. Server builders design machines around it, and this is ours.
  • Can it be configured?
    Yes. The processors, memory, storage and networking are set in your quote. If you would rather have eight separate GPUs on PCIe, with no NVLink between them, look at the EMARQUE AI Server.
  • What does the room need?
    An eight-GPU Blackwell Ultra machine is a data-centre load, not an office one: three-phase power and matching cooling. We survey power, cooling, rack depth and network drops before specifying anything.
  • What is the warranty?
    Three years from us on the system, with the component makers' warranties on top. We handle claims in Malaysia.
  • How does it compare to NVIDIA DGX B300?
    DGX B300 is NVIDIA's own machine on the same GPU board, sold as a fixed configuration with NVIDIA's software and support contract. This one is specified around your workload and supported by us here.
Next size upEMARQUE AI HGX RubinEight NVIDIA Rubin GPUs on one HGX Rubin NVL8 board, with 2 TB of GPU memory between them, liquid-cooled. Built to order for training large models, or serving them to a whole organisation.

Get a quote for the AI HGX B300.

Tell us what it needs to run and who will use it. We reply with a written quote in MYR.

Get a quote