Skip to content
EMARQUE AI

NVIDIA

NVIDIA DGX B300

Eight NVIDIA Blackwell Ultra GPUs in one air-cooled 10U system, with 2.1 TB of HBM3e. For reasoning models and inference at scale.
  • 8 × NVIDIA Blackwell Ultra GPUs, 2.1 TB HBM3e
  • 5th-gen NVLink with NVSwitch, 1.8 TB/s per GPU
  • 2 × Intel Xeon 6776P host
  • 8 × ConnectX-8 up to 800 Gb/s, 2 × BlueField-3 DPU
  • 30.7 TB NVMe E1.S for data
  • 10U, air-cooled, about 14 kW

The price depends on the networking and storage you choose, which maker's version you buy, and the software licences you add.

NVIDIA DGX B300

Key features

NVIDIA HGX B300 board

2.1 TB of GPU memory

Eight Blackwell Ultra GPUs on one NVLink domain hold half again the memory of NVIDIA DGX B200. That headroom is what long contexts and large key-value caches consume.
NVIDIA DGX B300 systems in a row of racks

Air-cooled, in a standard rack

NVIDIA designed the chassis for enterprise racks as well as MGX, so a data hall does not need liquid cooling to take one. It is a new chassis, not the DGX B200 one.
  • 800 Gb/s per adapter

    Eight ConnectX-8 adapters carry InfiniBand or Ethernet at twice the per-adapter rate of DGX B200, with two BlueField-3 DPUs taking storage and management traffic.

  • Set up and supported here

    Delivered and set up by us in Malaysia. We stay your contact here for the NVIDIA support contract.

Two ways to buy

NVIDIA's eight-GPU Blackwell Ultra platform reaches you two ways: NVIDIA's own DGX B300 system, or an HGX B300 server from an OEM that you configure, liquid-cooled builds included. We can source either.

NVIDIA

NVIDIA DGX B300

NVIDIA's own system in NVIDIA's 10U air-cooled chassis, with DGX OS and a three-year NVIDIA support contract. The configuration is NVIDIA's throughout.

  • NVIDIA's own 10U chassis
  • NVIDIA DGX OS included, with RHEL, Rocky and Ubuntu supported
  • Three-year NVIDIA business-standard support
OEM HGX B300 servers

HGX B300 from an OEM

NVIDIA HGX B300 in the OEM's own chassis, with CPU, memory, storage, networking and cooling picked from that vendor's list. Warranty and support come from the OEM.

  • Eight Blackwell Ultra GPUs on one NVLink domain
  • Air-cooled or direct-liquid-cooled chassis options
  • The OEM's warranty, which we handle in Malaysia
Our HGX B300 build: EMARQUE AI HGX B300⁠

Use cases

  • Reasoning

    Long contexts and multi-step work

    Models that answer in steps read and rewrite a long context on every pass. The GPU memory holds it, and NVIDIA's attention figure is the part of the machine that carries it.

  • Inference

    Many users from one machine

    Serving at FP4 or FP8 across eight GPUs, with enough memory to keep several models resident rather than loading them on demand.

  • Training

    Fine-tuning inside one node

    Full-parameter fine-tuning of large models across the NVLink domain, with no network between the GPUs holding the weights.

  • Cluster

    One node now, more later

    NVIDIA publishes DGX SuperPOD designs built from DGX B300 systems, so a second and third node join the same fabric as the first.

Specifications

NVIDIA DGX B300
GPU
8 × NVIDIA Blackwell Ultra SXM · 2.1 TB HBM3e total
GPU interconnect
5th-gen NVLink with NVSwitch · 1.8 TB/s per GPU
AI compute
FP4 144 PFLOPS with sparsity, 108 dense · FP8 72 PFLOPS with sparsity (NVIDIA figures)
vs NVIDIA DGX B200
1.5× the dense FP4 and 2× the attention performance (NVIDIA figure)
CPU
2 × Intel Xeon 6776P
Networking
8 × ConnectX-8 up to 800 Gb/s InfiniBand or Ethernet · 2 × BlueField-3 DPU
Management
1 GbE onboard NIC and 1 GbE host BMC, both RJ45
Storage
2 × 1.9 TB NVMe M.2 for the OS · 8 × 3.84 TB NVMe E1.S (30.7 TB)
Power
About 14 kW (NVIDIA; 14.5 kW in its user guide)
Form factor
10U rackmount, air-cooled
Software
NVIDIA DGX OS included, with RHEL, Rocky and Ubuntu supported. NVIDIA lists AI Enterprise and Mission Control as recommended, not included
Support
Three-year NVIDIA business-standard support

Questions about the DGX B300

  • How much does an NVIDIA DGX B300 cost in Malaysia?
    It depends on the networking and storage you choose, whether you buy NVIDIA's DGX B300 or an OEM's HGX B300 server, and the software licences you add. We quote it in writing, in MYR.
  • How is DGX B300 different from DGX B200?
    Same eight-GPU NVLink domain, newer GPUs. The Blackwell Ultra GPUs bring 2.1 TB of HBM3e against 1,440 GB, networking moves from ConnectX-7 at 400 Gb/s to ConnectX-8 at 800 Gb/s, and NVIDIA publishes 1.5 times the dense FP4 compute and twice the attention performance. The chassis is a new design, so plan the rack and power rather than assuming the DGX B200 footprint carries over.
  • Can it be configured?
    Not on DGX B300, where NVIDIA sets the CPU, memory and storage. If you need a say in those, or want liquid cooling, the OEM HGX B300 route is the one to quote.
  • Does it need liquid cooling?
    No. DGX B300 is air-cooled in a 10U chassis drawing about 14 kW, which a rack with the airflow and power for it can take. Direct-liquid-cooled HGX B300 servers are available from the OEMs if your rack density calls for them.
  • Is NVIDIA AI Enterprise included?
    No. DGX OS is included, and RHEL, Rocky and Ubuntu are supported. NVIDIA lists AI Enterprise and Mission Control as recommended, and both are licensed separately. We can quote them with the system.
  • What is the warranty, and who handles it here?
    DGX B300 carries a three-year NVIDIA business-standard support contract, and an OEM HGX system carries that vendor's warranty instead. Either way we are your contact in Malaysia and we lodge the claim.

Get a quote for the DGX B300.

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