Skip to content
EMARQUE AI

NVIDIA

NVIDIA DGX Spark

NVIDIA's GB10 Grace Blackwell Superchip in a desk-side computer with 128 GB of unified memory. We offer seven builds and set each one up here.
  • NVIDIA GB10 Grace Blackwell Superchip
  • 128 GB coherent unified memory, 273 GB/s
  • Up to 1 PFLOP FP4 with sparsity (NVIDIA figure)
  • ConnectX-7 at 200 Gb/s, 10 GbE and Wi-Fi 7
  • Up to 4 TB NVMe, by OEM build
  • 240 W external supply, from a wall socket

Delivered and set up by us in Malaysia. Prices for each variant are on emarque.co.

The DGX Spark family: NVIDIA's Founders Edition and the OEM builds

What the DGX Spark gives you

DGX Spark Founders Edition beside a laptop

The NVIDIA AI software stack, preinstalled

DGX OS, CUDA and the container runtime come installed, and NIM microservices download from NVIDIA's catalogue. NVIDIA AI Enterprise is a separate licence, which we can quote with the machine.
Two MSI EdgeXpert units cabled together over ConnectX-7

Link units for larger models

Link units over ConnectX-7: NVIDIA states two run models up to 400 billion parameters between them, and up to four can be connected. We set up the link if you buy more than one.
Rear ports of the MSI EdgeXpert build: USB-C, HDMI, 10 GbE and ConnectX-7

Desk-side, on office power

A 240 W external supply from a normal wall socket. ConnectX-7 at 200 Gb/s, 10 GbE and Wi-Fi 7 for the network.
  • Fine-tuning up to 70B parameters

    NVIDIA's figure for this machine. Past that, the work belongs on a multi-GPU server.

Is the DGX Spark right for you?

Developers, researchers and small teams who want NVIDIA's own desk-side system, running the NVIDIA AI software stack it ships with.

Choose it when

  • You want to develop against the software stack NVIDIA ships on its data-centre systems.
  • Your models fit in 128 GB of unified memory: NVIDIA states up to 200 billion parameters on one unit, and 400 billion with two linked.
  • A branch or edge location where shipping a rack is not realistic.

Choose something else when

  • A department has to be served from it at once. A multi-GPU server fits that better.
  • Fine-tuning goes beyond 70 billion parameters, which is NVIDIA's stated limit for this machine.
  • Most of the work is moving large datasets off disk, not holding a model in memory.

When you outgrow it

Step up to the AI PRO 500 for four GPUs in a rack, or to the NVIDIA DGX Station when the work needs its 748 GB of coherent memory on a desk.

What DGX Spark is for

DGX Spark Founders Edition at a developer's desk
Development

Model work on your own machine

Build and test prompts, retrieval pipelines and agents in the stack you would run on a server.
ASUS Ascent GX10 on an office desk beside a monitor
Private search

Chat and document search for one team

Index your documents and answer questions over them on the unit, so the files stay on it.
Two MSI EdgeXpert units in a 19-inch rack tray
On site

Inference where the data is

A factory floor, a branch office or a site where data cannot leave. It ships in one box.
DGX Spark in front of a DGX Station
Teaching

Teaching and lab work

Hands-on courses and student projects, without the lab buying a multi-GPU server.

Common across every OEM variant

The NVIDIA GB10 Grace Blackwell Superchip package, common to every build
Processor
NVIDIA GB10 Grace Blackwell Superchip
CPU
20-core Arm: 10 × Cortex-X925 + 10 × Cortex-A725
Memory
128 GB coherent unified LPDDR5x · 273 GB/s
AI compute
Up to 1 PFLOP FP4 with sparsity (NVIDIA figure)
Model size
Up to 200B parameters on one 128 GB unit, 400B on two linked; up to four can be linked (NVIDIA)
Fine-tuning
Up to 70B parameters (NVIDIA figure)
Networking
ConnectX-7 at 200 Gb/s · 10 GbE · Wi-Fi 7
Storage
Up to 4 TB NVMe, by OEM SKU
Power
240 W external supply (140 W GB10 TDP)
Software
NVIDIA AI software stack preinstalled. NVIDIA AI Enterprise is licensed separately
Warranty
The OEM's warranty, one or three years by variant. We handle claims in Malaysia

Questions about the DGX Spark

  • What does a DGX Spark cost in Malaysia?
    It depends on the build you choose, its storage and its warranty length. We quote in MYR, and per-build pricing sits on emarque.co.
  • Which OEM variant should I choose?
    The builds we offer share the chip and the 128 GB of memory, so the choice is storage and warranty length. Tell us what you plan to run and we will name one.
  • Who handles a warranty claim in Malaysia?
    We do. We lodge the claim with the maker and follow it for you.
  • Does it need special power or cooling?
    No. It runs from a standard wall socket and sits on a desk.
  • Can we order more than one unit?
    Yes. NVIDIA states that two linked units run models of up to 400 billion parameters, and that up to four can be linked. We set up the ConnectX link if you want them working together.

EMARQUE ATLAS is included

Ask questions and read documents with open models, on the Spark itself. Included with the DGX Spark systems we deliver; the download link follows your purchase.

See how ATLAS works⁠
EMARQUE ATLAS answering a question on a DGX Spark

Get your DGX Spark.

Tell us which variant you want, or what you want to run and we will recommend one.