The DGX Spark, measured.
Everything NVIDIA's spec sheet says, plus what a fleet operator would tell you about the parts it leaves out. Full specs, what actually fits in 128 GB, and where the machine loses.
Full specifications
Per node. Every GPUwerk instance is one whole machine, these numbers are all yours.
- Superchip
- GB10 Grace Blackwell
- AI performance
- 1 PFLOP (sparse FP4)
- Memory
- 128 GB unified LPDDR5x
- Memory bandwidth
- 273 GB/s
- CPU
- 20-core Arm (10× X925 + 10× A725)
- Storage
- 1 TB NVMe
- Network
- ConnectX-7 · 200 Gb/s
- Power
- 240 W max
- OS
- DGX OS 7 · Ubuntu 24.04
- CUDA
- Full stack preinstalled
273 GB/s memory bandwidth is the constraint that matters for single-stream LLM generation speed. The Spark's strength is capacity: models that simply don't fit on consumer GPUs, plus FP4 compute for batch serving and fine-tuning, where throughput scales far better than the single-stream numbers suggest.
Fleet benchmarks
We would rather show you an empty table than someone else's numbers. Our first full benchmark run is still going on the fleet, and results land here with methodology and raw logs when it finishes. Until then, the published third-party figures we trust are collected and argued with in the buying guide.
| Model | Quantization | Framework | Prefill tok/s | Generate tok/s | Max context |
|---|---|---|---|---|---|
| Qwen3 32B | Q4 | vLLM | measuring… | measuring… | measuring… |
| Llama 3.3 70B | Q4 | vLLM | measuring… | measuring… | measuring… |
| DeepSeek-R1 70B distill | Q4 | Ollama | measuring… | measuring… | measuring… |
Want to be notified when the numbers land, or run your own? Your first $100 top-up comes with $50 free, plenty for an evaluation day on a dedicated node.
What should I run on it?
Opinionated guidance from operating a fleet of these.
Team coding assistant
Qwen3-Coder 32B at Q4 leaves room for long contexts and several concurrent users. Serve with vLLM behind an OpenAI-compatible endpoint and point your editors at it.
Private chat + RAG
A 70B-class instruct model at Q4 fits in one node with headroom for an embedding model and a vector store on the same box, a whole private assistant on one machine. See the private ChatGPT setup →
Fine-tuning
LoRA fine-tunes of 70B-class models fit in unified memory without model-parallel gymnastics. Slower per step than a datacenter GPU, but it runs where consumer cards simply can't.
How it compares
Verdict-first comparisons against the machines buyers actually cross-shop. We say where the rival wins.
DGX Spark vs Mac Studio
Unified memory both ways, CUDA ecosystem vs raw bandwidth. Updated for Apple's M5 Ultra Mac Studio: the Mac wins decode, the Spark wins prefill and CUDA portability.
DGX Spark vs RTX 5090
32 GB of fast VRAM vs 128 GB of capacity. Not a close fight on speed, the whole question is whether your model fits.
DGX Spark vs Strix Halo
The other unified-memory contender, software stack decides it.
Guides & analysis
For anyone evaluating, buying, or setting up a DGX Spark, written from the fleet room, not a press kit.
DGX Spark setup: unboxing to first token
DGX OS first boot, SSH, drivers, and serving your first model, the guide we wish had existed when we racked our first unit.
What to know before buying a DGX Spark
Power, noise, networking two units, what it's genuinely good at, and the total cost of ownership against a 5090 rig or the cloud.
Serve Qwen3 with vLLM on a Spark
Copy-paste path to an OpenAI-compatible endpoint on your own hardware, Open WebUI frontend and systemd included.
The best models you're not allowed to send your data to
Which large open models fit in 128 GB / 256 GB, and how far behind the frontier they really are.
Fine-tune a model on your company's data
LoRA on 70B-class models fits in 128 GB unified memory, with the training set staying on the node's local NVMe.
An AI employee needs an office. Give it one you own.
OpenClaw, Hermes, and why an agent that reads your inbox belongs on hardware you control.
Automate the boring 80%
Batch document processing on a flat-rate machine, the marginal token is free, so run the model over everything.
Stop reading benchmarks. Run your own.
Deploy a dedicated Spark in under a minute, your first $100 top-up comes with $50 free, enough for a full evaluation day.
Deploy a Spark