Learn · NVIDIA Hardware & DGX · The DGX Family: Spark, Station
DGX Spark: The Mac-Mini-Sized Supercomputer
In plain English
DGX Spark is a tiny box — about the size of a Mac mini — that fits on your desk and runs huge AIs without the internet. Its brain, the GB10 chip, glues a CPU and GPU together so tightly they share one 128GB memory pool. Two Sparks can be chained with a network cable to run truly enormous models together! 🔗
What is actually going on
The DGX Spark packs the NVIDIA GB10 Grace Blackwell Superchip: a 20-core Armv9 CPU and a Blackwell GPU with 5th-gen Tensor Cores, joined by NVLink-C2C, with 128GB of coherent unified LPDDR5x memory. It delivers up to ~1 PFLOP FP4 (sparse) in a palm-sized, ~240W form factor and runs NVIDIA's full AI stack (CUDA, DGX OS, NIM). Two Sparks link over ConnectX networking (200GbE) to pool 256GB and serve models up to 405B parameters. Throughput is datacenter-adjacent, not RTX-class: capacity and software completeness are the pitch.
Why it matters
Spark brings the NVIDIA software ecosystem — the same stack as cloud datacenters — to a $3-5k desk box with 128GB of model capacity. Prototype and fine-tune locally, deploy the same artifacts to cloud DGX. For NVFP4-quantized models it's the sweet spot between Mac and server.
Another way to think about it
A tiny astronaut helmet that happens to contain a datacenter. It fits in a backpack, speaks fluent CUDA, and high-fives its twin over a cable to share brains. 🚀
Check yourself
How much unified memory does a single DGX Spark carry?
- 24 GB
- 64 GB
- 128 GB
- 748 GB
Show answer
128 GB — 128GB coherent unified memory shared by the Grace-architecture CPU and Blackwell GPU via NVLink-C2C.
How do two DGX Sparks run models up to 405B parameters?
- Stacked physically
- Linked over ConnectX high-speed networking to pool memory
- Via Bluetooth
- They can't
Show answer
Linked over ConnectX high-speed networking to pool memory — A ~200GbE ConnectX link lets two units act as one 256GB pool, splitting the model across them (tensor parallelism).
DGX Spark runs the same CUDA software stack used by NVIDIA cloud datacenters.
- True
- False
Show answer
True — That's the headline: develop locally on Spark, deploy identically to cloud DGX — no translation layer.