DGX Spark
A GB10 Grace Blackwell desktop AI computer with 128 GB of unified memory. These are the field notes for living with one.
Find your lane
Section titled “Find your lane” Tutorial → Brand-new box on the desk? Go from first boot to a model answering you, in under an hour.
How-to → Task in hand: network access, serving a local LLM, or chaining two and three Sparks.
Reference → Just the facts: full specs, the playbook index, the NCCL build, and the official docs map.
Explanation → The why behind the box: unified memory, FP4 and Blackwell, and multi-Spark networking.
How these notes are organized
Section titled “How these notes are organized”Structured by Diátaxis, so the kind of help you need maps to where it lives:
Directorytutorial/ learning-oriented
- first-boot
Directoryhow-to/ task recipes
- network-access
- run-a-local-llm
- run-vllm
- connect-two-sparks
- connect-three-sparks
- recover-the-system
Directoryreference/ dry lookup
- specifications
- playbooks
- nccl-build
- official-docs
Directoryexplanation/ the why
- unified-memory
- fp4-and-blackwell
- multi-spark-networking
The one thing to know before you plug it in
Section titled “The one thing to know before you plug it in”The DGX Spark powers on the instant power is applied. Attach your display, keyboard, mouse, and network cable before you connect the power supply, or you will boot headless into first-time setup with nothing attached.
Device photography is of the NVIDIA DGX Spark (Founders Edition). Product image via NVIDIA / Amazon listing; multi-unit photos via FiberMall and ServeTheHome. Used for identification and reference.
