Minisforum
Minisforum MS-A2 (Ryzen 9 9955HX)
AMD Ryzen 9 9955HX (16 cores, 32 threads) · 10 GbE · 1.78 L
8.4
$799Nov 92 videos
NVIDIA DGX Spark
NVIDIA's compact GB10 Grace Blackwell box brings 128GB of unified memory and 200 Gb/s ConnectX-7 networking to a desk for local AI work.
- NVIDIA GB10 Grace Blackwell Superchip (20-core Arm: 10 Cortex-X925 + 10 Cortex-A725)
- 128 GB
- 4,000 GB
- 10 GbE
PCPR score
7.1
Fair
Best price we’ve seen
See retailers
Check retailers for today’s price.
R1
The short version
The DGX Spark is NVIDIA's compact desktop for AI development, built on the GB10 Grace Blackwell Superchip with 128GB of unified LPDDR5X memory, a 10GbE port and a ConnectX-7 200 Gb/s NIC, running NVIDIA DGX OS. Tom's Hardware found it a fast, well-documented local AI machine that can beat Ryzen AI Max+ 395 systems, but expensive if you do not use its clustering and CUDA strengths, and it does not run Windows. Treat it as an AI development appliance rather than a general desktop.
Pros
- 128GB of coherent unified LPDDR5X memory for running large local AI models
- ConnectX-7 NIC at up to 200 Gb/s plus 10GbE Ethernet and Wi-Fi 7
- Up to four systems can be clustered over ConnectX-7, per NVIDIA
- Tom's Hardware found the software stack polished and well documented
- Compact 150 x 150 x 50.5 mm chassis weighing 1.2 kg
Cons
- Tom's Hardware says it is expensive if its AI and clustering features go unused
- Does not run Windows; it ships with NVIDIA's Linux-based DGX OS
- Memory bandwidth limits it against big GPUs for some AI work, per ServeTheHome
- Powered through a USB-C input rather than a standard internal supply
Best for
- Local AI and large-model development
- Researchers who want NVIDIA CUDA and DGX tooling on a desk
J3
Full specifications
Core
- Processor
- NVIDIA GB10 Grace Blackwell Superchip (20-core Arm: 10 Cortex-X925 + 10 Cortex-A725)
- Graphics
- NVIDIA Blackwell (integrated in GB10)
Memory
- Memory
- 128 GB
- Memory type
- LPDDR5X coherent unified memory
- Max memory
- 128 GB
- Storage
- 4,000 GB
Connectivity
- Wi-Fi
- Wi-Fi 7
- Ethernet
- 10 GbE
Ports & I/O
- HDMI outputs
- 1
- DisplayPort outputs
- 0
Physical
- Weight
- 1.2 kg
Software
- Operating system
- Linux
Compatibility notes
- Ships with NVIDIA DGX OS, a Linux distribution based on Ubuntu, on the Arm-based GB10 platform; it does not run Windows
- 128GB of LPDDR5X unified memory; a 64GB configuration is available only through OEM partners
- Storage is M.2 NVMe up to 4TB with self-encryption; Tom's Hardware notes the SSD is a user-replaceable M.2 2242 drive
- Powered through a USB-C input from a 240 W supply; body measures 150 x 150 x 50.5 mm and weighs 1.2 kg
- Two QSFP ports connect the ConnectX-7 NIC for clustering up to four systems
P1
Where to buy
- Buy from the maker
NVIDIA official page
Direct link
- Check price on Amazon
Amazon
Search this retailer
- See it at Newegg
Newegg
Search this retailer
- See it at Best Buy
Best Buy
Search this retailer
- See it at B&H
B&H Photo
Search this retailer
- See it at Micro Center
Micro Center
Search this retailer
- Search TikTok
TikTok Shop
Search this retailer
We may earn a commission when you buy through links on this page. It never affects our scores or what we recommend. Prices move quickly, so confirm the final price at the retailer.
AV1
Video reviews
Overview · Comparison
M5 Ultra vs 2 DGX Sparks… The Number You're Not Looking At
Alex Ziskind
Compares an Apple M5 Ultra machine with a pair of clustered DGX Spark units for local AI workloads.
Overview · Review
NVIDIA DGX Spark – A Non-Sponsored Review (Strix Halo Comparison, Pros & Cons)
Bijan Bowen
An unsponsored review of the NVIDIA DGX Spark covering unboxing, first impressions, AI workload testing, and comparisons against Strix Halo.
Positive · Review
128GB with 200GbE NVIDIA DGX Spark is GREAT for Local AI
ServeTheHome
ServeTheHome reviews the NVIDIA DGX Spark, covering its 128GB memory, Blackwell GPU, 200GbE networking, and local AI use.
Overview · Review
Deep Dive into Nvidia's DGX Spark GB10
Level1Techs
Level1Techs unboxes and deep dives into the DGX Spark with demos of AI workloads, fine-tuning, and networking.
AV2
What reviewers and owners say
Tom's Hardware
PositiveFound the GB10 system quick and well supported for local AI thanks to 128GB of memory and NVIDIA's software stack, but costly unless its clustering features are used, and it lacks Windows support.
Review
Read the sourceServeTheHome
PositiveSaw the GB10 box as a significant step for running large local AI models with 128GB of unified memory, while cautioning that the platform still felt immature at launch and that memory bandwidth limits it against big GPUs.
Review
Read the sourceClub386
MixedCollected early reviews praising its memory capacity, clustering support and developer-ready software for local AI work, while cautioning that its strong inference performance comes at a high price.
2025-10-15
Read the source
Summaries are written in our words. Follow the links for each source’s own review.
R9
Common questions
Does the DGX Spark run Windows?
No. It ships with NVIDIA's DGX OS, a Linux distribution based on Ubuntu, and Tom's Hardware notes it does not run Windows yet. It is aimed at AI development rather than general desktop use.
Can several DGX Spark systems be linked together?
Yes. NVIDIA says up to four systems can be connected over ConnectX-7 networking, with direct QSFP cables for two or three systems and a 200GbE switch required for four.
How much memory does it have, and can it be expanded?
NVIDIA lists 128GB of LPDDR5X unified memory, with a 64GB configuration available through OEM partners. No memory upgrade path is described, though Tom's Hardware says the M.2 SSD is user-replaceable.
J9Alternatives
