Skip to content
EMARQUE AI

NVIDIA

NVIDIA DGX B200

Eight NVIDIA Blackwell GPUs linked by NVLink in one air-cooled 10U system, with 1,440 GB of HBM3e. For training and inference on your own premises.
  • 8 × NVIDIA Blackwell GPUs, 1,440 GB HBM3e
  • 5th-gen NVLink with 2 × NVSwitch
  • 2 × Intel Xeon Platinum 8570, 2 TB system memory, configurable to 4 TB
  • 8 × ConnectX-7 up to 400 Gb/s, 2 × BlueField-3 DPU
  • 30.7 TB NVMe for data, plus 2 × 1.9 TB for the OS
  • 10U, air-cooled, up to 14.3 kW

The price depends on the system memory, whether you buy NVIDIA's DGX B200 or an OEM's HGX B200 server, and the software licences you add.

NVIDIA DGX B200

Key features

NVIDIA DGX B200, exploded view

Eight GPUs, one memory pool

NVLink and the two NVSwitch chips link all eight GPUs at 14.4 TB/s aggregate, with 1,440 GB of HBM3e between them. A model too large for one GPU runs across the node instead of across machines.
NVIDIA DGX B200

A host sized for the GPUs

Two Xeon Platinum processors, system memory up to 4 TB and 30.7 TB of internal NVMe keep data moving to the GPUs. The configuration is NVIDIA's, so there are no CPU options.
  • Networking for a second node

    Eight ConnectX-7 adapters carry InfiniBand or Ethernet, and two BlueField-3 DPUs carry storage and in-band management traffic. Those ports are what another node connects to.

  • 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

Eight B200 GPUs with 1,440 GB of HBM3e and NVLink at 14.4 TB/s (NVIDIA figures) reach you two ways: NVIDIA's own DGX B200 system, or an OEM HGX B200 server you configure. We can source either.

NVIDIA

NVIDIA DGX B200

NVIDIA's own system in NVIDIA's 10U chassis, with DGX OS and a three-year NVIDIA support contract. The configuration is fixed apart from the system memory.

  • NVIDIA's own 10U air-cooled chassis
  • NVIDIA DGX OS and three-year Enterprise Business-Standard Support
  • System memory 2 TB or 4 TB, no CPU or chassis choices
OEM HGX B200 servers

HGX B200 from an OEM

An NVIDIA HGX B200 platform in an OEM server maker's own chassis, with CPU, memory, storage and networking picked from that maker's list. Warranty and support come from the OEM.

  • Same GPU count, GPU memory and NVLink bandwidth (NVIDIA figures)
  • CPU, memory, NVMe layout and networking are yours to choose
  • The OEM's warranty, which we handle in Malaysia

Use cases

  • Training

    Large models trained in one node

    Pre-training and full-parameter fine-tuning where every GPU works on the same weights, with the NVLink fabric carrying the traffic instead of a network.

  • Inference

    Serving a whole organisation

    Several models, or several versions of one, served from the same node with the GPU memory to hold long contexts.

  • Cluster

    One node now, more later

    NVIDIA publishes DGX SuperPOD designs built from DGX B200 systems. The ConnectX-7 ports are the fabric a second and third node join.

Specifications

NVIDIA DGX B200
GPU
8 × NVIDIA Blackwell · 1,440 GB HBM3e at 64 TB/s
GPU interconnect
5th-gen NVLink, 2 × NVSwitch · 14.4 TB/s aggregate
AI compute
FP4 144 PFLOPS with sparsity, 72 dense · FP8 72 PFLOPS with sparsity, 36 dense (NVIDIA figures)
CPU
2 × Intel Xeon Platinum 8570 · 112 cores · 2.1 GHz base, 4 GHz max boost
System memory
2 TB in 32 DIMMs, configurable to 4 TB
Networking
8 × ConnectX-7 up to 400 Gb/s InfiniBand or Ethernet · 2 × BlueField-3 DPU
Management network
10 Gb/s onboard NIC (RJ45) · 100 Gb/s dual-port Ethernet NIC · BMC (RJ45)
Storage
OS on 2 × 1.9 TB NVMe M.2 (RAID 1) · data on 8 × 3.84 TB self-encrypting NVMe U.2 (30.7 TB, RAID 0)
Power
About 14.3 kW max · 6 × 3.3 kW supplies, 5+1 redundant (NVIDIA figures)
Cooling
Air-cooled, 1,550 CFM · 10 to 35 °C
Form factor
10U rackmount · 444 × 482 × 897 mm · up to 142.4 kg
Software
DGX OS or Ubuntu included. NVIDIA lists AI Enterprise and Mission Control as recommended, not included
Support
Three-year NVIDIA Enterprise Business-Standard Support

Questions about the DGX B200

  • What is the difference between DGX B200 and an HGX B200 server?
    Both have eight B200 GPUs with the same GPU memory and NVLink bandwidth (NVIDIA figures). DGX B200 is NVIDIA's own system, in NVIDIA's chassis, with DGX OS and NVIDIA's support contract, in a configuration you do not change. An HGX B200 server is built by an OEM server maker on NVIDIA's HGX B200 platform, so the chassis, CPU, memory, storage and networking come from that maker's list, and so does the warranty.
  • Can it be configured?
    On DGX B200, only the system memory, which is 2 TB or 4 TB. If you need a say in the CPU, the storage layout or the networking, the OEM HGX B200 route is the one to quote.
  • Is NVIDIA AI Enterprise included?
    No. DGX OS or Ubuntu is included. NVIDIA lists AI Enterprise and Mission Control as recommended, not included. We can quote them with the system so the software is ready when the machine arrives.
  • What is the warranty, and who handles it here?
    DGX B200 comes with NVIDIA's three-year Enterprise Business-Standard Support for hardware and software; an OEM HGX B200 server carries that maker's warranty. Either way it is the maker's warranty, handled by us in Malaysia: we lodge the claim and follow it for you.
  • What if we need more than this?
    The DGX B300 is NVIDIA's eight-GPU Blackwell Ultra system, with more GPU memory (2.1 TB) and faster networking. Above that, the NVIDIA GB300 NVL72 is rack-scale, which we source through NVIDIA's OEM partners and coordinate the install for.

Get a quote for the DGX B200.

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