Moving up a class
NVIDIA DGX Spark vs NVIDIA DGX Station
Memory, bandwidth, software stack, price and the models each one can hold. Generated from the two sourced machine records; estimates are labelled.
NVIDIA GB10
NVIDIA DGX Spark
128 GB unified
Open the record
NVIDIA GB300
NVIDIA DGX Station
496 GB CPU + 252 GB GPU · 748 GB coherent
Open the recordWhat actually differs
- Different software stacks: the NVIDIA DGX Spark runs CUDA, the NVIDIA DGX Station runs CUDA. Check that your runtime and model format are supported before comparing numbers.
- The NVIDIA DGX Station has 124 GB more model memory (252 GB vs 128 GB). That decides which models fit at all.
- The NVIDIA DGX Station reads memory 26.0× faster (7100 vs 273 GB/s). For a model that fits on both, that is the ceiling on generation speed.
- No public list price in the record for the NVIDIA DGX Station. Ask the supplier for a configured quote.
- Form factor: Desktop AI system vs Deskside workstation.
Which models fit
Estimate · 4-bit · 8K context · 1 request| Model | NVIDIA DGX Spark | NVIDIA DGX Station |
|---|---|---|
| Llama 3.1 8B | Fits 16 GB≤ 48 tok/s | Fits 12 GB≤ 1251 tok/s |
| Qwen3 14B | Fits 20 GB≤ 28 tok/s | Fits 16 GB≤ 721 tok/s |
| Qwen3 32B | Fits 32 GB≤ 13 tok/s | Fits 28 GB≤ 338 tok/s |
| Llama 3.3 70B | Fits 54 GB≤ 6.3 tok/s | Fits 50 GB≤ 164 tok/s |
| gpt-oss-20b (MoE) | Fits 23 GB≤ 110 tok/s | Fits 19 GB≤ 2871 tok/s |
| gpt-oss-120b (MoE) | Fits 80 GB≤ 77 tok/s | Fits 76 GB≤ 2008 tok/s |
| Qwen3 235B-A22B (MoE) | Too large 151 GB | Fits 147 GB≤ 499 tok/s |
| DeepSeek V3 / R1 671B (MoE, MLA) | Too large 408 GB | Too large 404 GB |
Specifications
Sourced and dated| NVIDIA DGX Spark | NVIDIA DGX Station | |
|---|---|---|
| Chip / accelerator | GB10 Grace Blackwell Superchip | GB300 Grace Blackwell Ultra Desktop Superchip |
| Platform | NVIDIA GB10 | NVIDIA GB300 |
| Model memory | 128 GB unified | 496 GB CPU + 252 GB GPU · 748 GB coherent |
| Memory bandwidth | 273 GB/s | GPU 7.1 TB/s · CPU 396 GB/s · NVLink-C2C 900 GB/s |
| Software stack | CUDA | CUDA |
| Operating system | NVIDIA DGX OS | NVIDIA DGX OS (Ubuntu) · Windows edition announced for Q4 2026 |
| Storage | 4 TB NVMe | 4 × M.2 PCIe Gen 5 slots · drives configured by the partner |
| Networking | 10 GbE · ConnectX-7 · Wi-Fi 7 | ConnectX-8 SuperNIC · 2 × 400 GbE QSFP112 · 10 GbE · 1 GbE BMC |
| Form factor | Desktop AI system | Deskside workstation |
| Dimensions | 150 × 150 × 50.5 mm | Set by each partner’s chassis |
| Weight | 1.2 kg | Set by each partner’s chassis |
| Power rating (not consumption) | 240 W adapter · 140 W SoC TDP | 1,600 W platform system-power specification · 20 A circuit specified by NVIDIA |
| Supplier list price | ≈ $4.7k · Supplier list price · United States · Sept 2026 | Price from the manufacturer · No public list price · manufacturer page |
| Availability | Listed for order · recheck with supplier | Order through partner manufacturers · configuration and price set by each partner |
| Model configurations in the Lab | 6 | 4 |
| Setups in the Lab | 5 | 3 |
| Record reviewed | 2026-09-13 | 2026-09-17 |
Before you choose
NVIDIA DGX Spark
- Memory capacity does not establish interactive speed or simultaneous request capacity.
- UT has documented a Spark deployment. Confirm exact software versions and support scope.
NVIDIA DGX Station
- 748 GB coherent memory is 252 GB HBM3e on the GPU plus 496 GB LPDDR5X on the CPU, joined by NVLink-C2C. The two tiers have very different bandwidths; large models spill from the fast tier into the slow one.
- NVIDIA does not sell the DGX Station directly: it is ordered through partner manufacturers, each with its own chassis, drives, warranty and price. Partner listings observed in 2026 ran from about $85,000 to $175,000 depending on configuration.
- NVIDIA specifies 1,600 W total system power and a 20 A circuit. Confirm OEM and regional installation requirements before choosing an office outlet. The GB300 can be paired with an RTX PRO 6000 Blackwell workstation GPU for graphics and additional compute.
NVIDIA DGX Spark compared with
ASUS Ascent GX10 Dell Pro Max with GB10 HP ZGX Nano G1n Lenovo ThinkStation PGX Acer Veriton GN100 GIGABYTE AI TOP ATOM MSI EdgeXpert Apple Mac Studio · M3 Ultra Apple Mac Studio · M5 Ultra Framework Desktop · Ryzen AI Max+ MINISFORUM MS-S1 MAX HP Z2 Mini G1a AMD Ryzen AI HaloOwn one of these? A measured run (model revision, runtime, context, tokens per second) is worth more than this whole page. Publish it.