A good fit
- CUDA developers who want the NVIDIA AI stack on their desk
- Fine-tuning experiments and prototypes
- Private AI for one to five people at the same time, with documents that stay in your office
Home / Workstations / NVIDIA DGX Spark
A 128 GB AI supercomputer the size of a book
The NVIDIA DGX Spark runs on the GB10 Grace Blackwell chip with 128 GB of unified memory. We supply it as the ASUS Ascent GX10 or the MSI EdgeXpert, with a 1 TB or 4 TB SSD.
Price on request
Tell us how many you need and we send a written quote.
128 GBunified memory
1 PFLOPAI performance, FP4 with sparsity
20-coreArm CPU on the same chip
2 × 128 GBcan be linked as one
What it runs, and how fast
The 128 GB of memory is the point: one box holds models of around 120 billion parameters that do not fit on a normal graphics card. Mixture-of-experts (MoE) models are the sweet spot, because only part of the model works on each word.
| Model | Speed on one box | In plain words |
|---|---|---|
| gpt-oss-120b | about 42 tok/s | Very fast, a large MoE model |
| Qwen3-Coder-Next | about 56 tok/s | Very fast, for coding |
| Gemma 4 26B-A4B | about 47 tok/s | Very fast, small MoE model |
| Qwen3.5-122B-A10B | about 26 tok/s | Fast, a large MoE model |
| Qwen3.8 27B | about 10 tok/s | About reading speed, a dense model |
Estimates from the EMD Lab AI selector: calculated, not measured. One user, Q4_K_M quantisation, 32K context. Dense 27–31B models are slower because every word uses the whole model, so pick MoE models where you can. For a measured comparison, the llama.cpp community's results on a DGX Spark show gpt-oss-120b writing at about 46–61 tok/s and reading a 2,048-token prompt at 1,737–1,956 tok/s (llama.cpp discussion 16578; community results, not EMD Lab measurements).
Other models and other hardware: Which hardware runs which AI?
Who it is for, and who it is not
Look at something bigger
- A team that reads long documents all day: the AI workstation or the QNAP AI server
- The very largest models: the ASUS ET900N G3 with 748 GB of memory
Choose your box
Both are the same NVIDIA DGX Spark inside. Pick the maker and the SSD; the memory is 128 GB on both.
ASUS Ascent GX10
- ChipNVIDIA GB10 Grace Blackwell
- Memory128 GB
- Storage1 TB
- PriceOn request
We confirm the exact configuration with you in the quote.
Two boxes, one bigger machine
Each box has a ConnectX-7 network port. Two 128 GB boxes joined with a cable work together as 256 GB, and NVIDIA says a linked pair handles models of up to 400 billion parameters, against 200 billion for a single box. Four boxes need an external 200 GbE switch. Ask us before you plan a cluster.
Specifications
Both boxes (NVIDIA DGX Spark)
- Superchip
- NVIDIA GB10 Grace Blackwell: Blackwell GPU with 5th-generation Tensor Cores and 4th-generation RT Cores, plus a 20-core Arm CPU (10 Cortex-X925 and 10 Cortex-A725) joined by NVLink-C2C
- AI performance
- Up to 1 petaFLOP at FP4, with sparsity
- Memory
- 128 GB LPDDR5X coherent unified memory, 256-bit, 273 GB/s
- Network
- 10 GbE RJ-45; NVIDIA ConnectX-7 SmartNIC (200 Gbps); Wi-Fi 7; Bluetooth 5.4
- Operating system
- NVIDIA DGX OS (Ubuntu-based Linux) with the NVIDIA AI software stack
The two boxes side by side
| ASUS Ascent GX10 | MSI EdgeXpert | |
|---|---|---|
| Storage | NVMe M.2 SSD; the maker lists 1, 2 and 4 TB. We offer 1 TB and 4 TB. | NVMe M.2 SSD with self-encryption; the maker lists 1, 2 and 4 TB for the 128 GB model. We offer 1 TB and 4 TB. |
| Ports | 3 × USB-C 20 Gbps with DisplayPort 2.1, 1 × USB-C power input, HDMI 2.1a, 10 GbE | 4 × USB-C up to 20 Gbps, HDMI 2.1, DisplayPort via USB-C, 10 GbE |
| Size | 150 × 150 × 51 mm | 151 × 151 × 52 mm |
| Weight | 1.48 kg | 1.2 kg |
| Power adapter | 240 W | Not stated on MSI's page (NVIDIA lists 240 W for the DGX Spark) |
| Cooling | Dual-fan cooling with dual vapor chamber | FanDock design with SSD vapor chamber |
Specifications from NVIDIA, ASUS and MSI, read 25 October 2026. Ask us about warranty and support.
Talk to us about the DGX Spark
Tell us what you want to run and how many people will use it. We answer with a written quote.