One developer, or a small team trying models
DGX Spark. 128 GB of unified memory runs models up to 200B parameters on one unit (NVIDIA figure), from a wall socket on a desk.
DGX Spark vs DGX Station
NVIDIA's two desk-side systems, both on Grace Blackwell: DGX Spark on the GB10 and DGX Station on the GB300. Every figure here is NVIDIA's own, with sparse and dense marked where NVIDIA marks them.
Published . Figures from NVIDIA's DGX Spark and DGX Station pages.
The two systems.
NVIDIA
NVIDIA DGX Spark
A desk-side AI computer on NVIDIA's GB10 Grace Blackwell Superchip, with 128 GB of unified memory. For developers and small teams running models on their own machine.
- Unified memory
- 128 GB
- FP4 with sparsity
- 1 PFLOP
- Parameters on one unit
- 200B
NVIDIA
NVIDIA DGX Station
A desk-side system on NVIDIA's GB300 Grace Blackwell Ultra Desktop Superchip, with 748 GB of coherent memory. For research and data science teams running large models on their own hardware.
- Coherent memory
- 748 GB
- FP4 with sparsity
- 20 PFLOPS
- ConnectX-8 networking
- 800 Gb/s
Desk-side systems
DGX Spark and DGX Station, side by side.
| NVIDIA figure | DGX Spark | DGX Station |
|---|---|---|
| Superchip | NVIDIA GB10 Grace Blackwell | NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip |
| Memory | 128 GB unified LPDDR5x | 748 GB coherent: 252 GB HBM3e plus 496 GB LPDDR5X |
| FP4 compute | Up to 1 PFLOP with sparsity | 20 PFLOPS with sparsity, 15 dense |
| Model size | Up to 200B parameters on one unit; 400B on two linked units | Up to about a trillion parameters on one unit, depending on quantisation |
| Sharing | One developer or a small team | Up to 7 MIG instances, isolated in hardware |
| Power | 240 W external supply, from a wall socket | 1,600 W; NVIDIA states a 20 A circuit |
| Software | DGX OS, CUDA and the DGX Dashboard preinstalled | Ubuntu with NVIDIA AI Developer Tools |
Sources: NVIDIA's DGX Spark and DGX Station pages and the DGX Spark hardware guide. The OEM builds share each platform's chip and memory.
Which one fits.
A group sharing one machine
DGX Station. MIG splits its GPU into as many as seven instances, isolated in hardware, so a team shares the machine without sharing a session.
Larger models, still at a desk
DGX Station. 748 GB of coherent memory holds models of up to about a trillion parameters on one unit, depending on quantisation (NVIDIA).
More than one Spark
Two linked DGX Spark units run models of up to 400B parameters, and up to four can be linked (NVIDIA figures). A Station is one machine with far more memory.
Common questions.
What is the difference between DGX Spark and DGX Station?
Size and memory. DGX Spark is NVIDIA's small desk-side system on the GB10 Superchip with 128 GB of unified memory. DGX Station is the larger desk-side system on the GB300 Desktop Superchip with 748 GB of coherent memory and 20 PFLOPS of FP4 with sparsity (NVIDIA figures).Is DGX Station faster than DGX Spark?
On NVIDIA's FP4 figures, yes: 20 PFLOPS with sparsity for the Station against up to 1 PFLOP for the Spark. The bigger difference for most teams is memory: 748 GB against 128 GB.What power does each need?
DGX Spark runs from a 240 W external supply on a normal wall socket. DGX Station draws 1,600 W and NVIDIA states a 20 A circuit, which a standard Malaysian 13 A socket is not, so we go through the room with you first.Can I buy both in Malaysia?
Yes. Both come in builds from NVIDIA's OEM partners, and DGX Spark also as NVIDIA's own Founders Edition. We quote them in writing, in MYR, then deliver, set up and support them in Malaysia. The emarque.co store lists the DGX Spark builds with prices.What comes after DGX Station?
A rackmount system. Our AI PRO 500 takes up to four NVIDIA RTX PRO 6000 Blackwell GPUs with 384 GB of GPU memory in a 5U chassis, and our AI Server takes six or eight.
The systems on this page.
Tell us about your workload.
Tell us what you want to run and who will use it. We recommend a system in writing, in MYR.