Skip to content
EMARQUE AI

NVIDIA

NVIDIA DGX Station GB300

NVIDIA's desk-side system on the GB300 Grace Blackwell Ultra Desktop Superchip, with 748 GB of coherent memory. We deliver it and set it up in Malaysia.
  • NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip
  • 748 GB coherent memory: 252 GB HBM3e and 496 GB LPDDR5X
  • 20 PFLOPS FP4 with sparsity, 15 dense (NVIDIA figures)
  • ConnectX-8 SuperNIC, up to 800 Gb/s Ethernet
  • Up to 7 MIG instances on the GPU
  • 1,600 W; NVIDIA states a 20 A circuit

The price depends on which OEM's build you choose and the configuration.

NVIDIA DGX Station, GIGABYTE W775 build, front three-quarter view

Product overview

Front of the GIGABYTE W775 build
Front
Rear ports of the GIGABYTE W775 build
Rear I/O, including the network ports
Inside the GIGABYTE W775 build
Inside, side panel off

What the GB300 platform gives you

The NVIDIA GB300 Desktop Superchip board, Grace CPU and Blackwell Ultra GPU together

One address space across CPU and GPU

The Grace CPU's memory and the GPU's HBM3e sit in one coherent 748 GB space over NVLink-C2C, so a model and its context are not copied between host and device.
NVIDIA DGX Station, GIGABYTE W775 build

Delivered and set up by us

We deliver the OEM build that suits you and set the unit up on your site, with support from us in Malaysia afterwards.
  • Models up to about a trillion parameters

    NVIDIA states that models of that size run on one unit, which keeps the model and the data on your own machine.

  • Up to seven instances on one GPU

    MIG divides the GPU into as many as seven instances, isolated in hardware, so a group shares the machine without sharing a session.

  • The software NVIDIA ships with it

    Ubuntu with the NVIDIA AI Developer Tools. NVIDIA AI Enterprise is a separate licence, which we can quote with the system.

Three builds of one platform

NVIDIA defines the platform and its OEM partners build the chassis, so the compute, memory, networking and software are the same in all three. We recommend whichever suits your software and your room. Price on request.

What we include with every DGX Station

  • Pre-sales consultation and sizing to the models you plan to run
  • Delivery and setup on your site
  • The maker's warranty, handled in Malaysia
  • Support from us in Malaysia after handover

What’s inside the GB300 platform

The GB300 Desktop Superchip board beside the NVIDIA DGX Station

Compute

  • NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip
  • 20 PFLOPS FP4 with sparsity, 15 PFLOPS dense (NVIDIA figures)
  • NVLink-C2C at 900 GB/s between CPU and GPU

Memory

  • 252 GB of HBM3e at 7.1 TB/s on the GPU
  • 496 GB of LPDDR5X at 396 GB/s on the CPU
  • 748 GB coherent: one address space across both

Networking

  • ConnectX-8 SuperNIC, up to 800 Gb/s Ethernet
  • 2 × QSFP112 at 400 Gb/s each
  • 10 GbE, and a 1 GbE port for management

Software and partitioning

  • Ubuntu with the NVIDIA AI Developer Tools
  • NVIDIA AI Enterprise licensed separately
  • Up to 7 MIG instances on the GPU

Is the DGX Station right for you?

Research labs and data-science teams who want NVIDIA's desk-side Blackwell Ultra system rather than a rackmount server.

Choose it when

  • You want NVIDIA's own desk-side system, running Ubuntu with the NVIDIA AI Developer Tools it ships with.
  • The work needs 748 GB of coherent memory in one chassis: 252 GB of HBM3e and 496 GB of LPDDR5X (NVIDIA figures).
  • The same work will later run on NVIDIA rack systems.

Choose something else when

  • The work runs on a DGX Spark or an AI PRO 500.
  • The growth path is rack-scale. Go to the NVIDIA DGX B200 or DGX B300, or a GB300 NVL72 rack.
  • The room cannot give it 1,600 W on a 20 A circuit, which is what NVIDIA states.

When you outgrow it

When a desk-side machine is not enough, move to the single-node NVIDIA DGX B200 or DGX B300 in 10U, or to a GB300 NVL72 rack. We source the rack-scale systems through NVIDIA's OEM partners and coordinate the install.

What DGX Station is for

  • Private models

    Models that answer from your own files

    Serve an open-weight chat or vision model from the Station, with the prompts and the documents staying in your building.

  • Fine-tuning

    Adapters trained on your own data

    Train a LoRA or QLoRA adapter on a domain dataset, then evaluate it on the same unit.

  • Before rack-scale

    A step before a rack system

    Build the workload here and measure it, then move it on when more people need it. We source rack systems through NVIDIA's OEM partners and coordinate the install.

Specifications

Rear ports of the GIGABYTE W775 build
GPU
1 × NVIDIA Blackwell Ultra · 252 GB HBM3e at 7.1 TB/s
CPU
1 × NVIDIA Grace, 72 Arm Neoverse V2 cores · 496 GB LPDDR5X at 396 GB/s
Coherent memory
748 GB (252 GB HBM3e + 496 GB LPDDR5X)
AI compute
20 PFLOPS FP4 with sparsity · 15 PFLOPS dense (NVIDIA figures)
Interconnect
NVLink-C2C at 900 GB/s between CPU and GPU
Networking
ConnectX-8 SuperNIC up to 800 Gb/s Ethernet · 2 × QSFP112 at 400 Gb/s · 10 GbE · 1 GbE management
MIG partitions
Up to 7 isolated GPU instances
Storage
4 × M.2 Gen 5 slots
Software
Ubuntu with NVIDIA AI Developer Tools. NVIDIA AI Enterprise is licensed separately
Power
1,600 W. NVIDIA states a 20 A circuit
Warranty
3 years from the OEM. We handle claims in Malaysia
OEM variants
MSI XpertStation WS300 · ASUS ExpertCenter Pro ET900N G3 · GIGABYTE W775-V10-L01

Common questions

  • What is the NVIDIA DGX Station?
    NVIDIA's desk-side system on the GB300 Grace Blackwell Ultra Desktop Superchip, where the Grace CPU and the Blackwell Ultra GPU share one coherent memory space. NVIDIA's OEM partners build the chassis; we deliver and set up the unit in Malaysia.
  • How large a model will it run?
    NVIDIA states that models up to about a trillion parameters run on one unit, which depends on the quantisation. Training a model that size is rack-scale work.
  • What power does it need?
    1,600 W, and NVIDIA states a 20 A circuit. A standard Malaysian 13 A wall socket is not a 20 A circuit, so we go through the socket and the room with you before the unit ships.
  • What is the warranty?
    Three years from the OEM. We handle the claim in Malaysia.
  • Can I buy a DGX Station in Malaysia?
    Yes. NVIDIA's OEM partners build it, and we quote the three builds shown on this page in writing, in MYR, then deliver, set up and support the unit in Malaysia.
  • How much does a DGX Station cost in Malaysia?
    It depends on which OEM's build you choose and its configuration. We quote it in writing, in MYR, and the emarque.co store lists the builds it carries.
  • How does it compare with DGX Spark?
    DGX Spark is the smaller desk-side system, on the GB10 Superchip with 128 GB of unified memory, for one developer or a small team. The Station holds more memory and splits into MIG instances for a group.

Talk to us about a DGX Station.

Tell us the models you want to run and where the unit will sit. We quote in writing.