NVIDIA's DGX Spark launches, a 128GB desktop AI machine
Originally announced by NVIDIA . We link the primary source so you can read it for yourself.
NVIDIA’s DGX Spark, a small desktop AI computer built on the GB10 Grace Blackwell chip with 128GB of unified memory, has gone on sale through NVIDIA’s own store and partners including Dell, Asus, MSI, and HP. It brings NVIDIA’s software ecosystem to a compact local machine aimed at developers and prototyping.
What it is
The appeal is the combination: 128GB of unified memory for holding large models, with CUDA compatibility so the same code and tooling used on data-centre GPUs runs locally. That makes it interesting for people who develop against NVIDIA’s stack and want to work on large models on their own desk before moving to bigger hardware.
The trade-off
As with other unified-memory machines, bandwidth is the limit. Memory bandwidth on the DGX Spark is modest next to a discrete GPU, so generation is slower than a card would be, and it is aimed more at development and prototyping than at maximum serving throughput. It also sits above unified-memory mini PCs on price, so the value case depends on how much you need NVIDIA’s ecosystem specifically.
Why it matters
The DGX Spark was an early entrant in a trend that has since grown: unified-memory machines that trade bandwidth for the capacity to hold large models at low power. It now sits alongside high-memory Macs and, more recently, Strix Halo mini PCs. For local AI, that widens the practical options beyond stacking discrete GPUs. Which one suits you depends on your software needs and your tolerance for slower generation, a trade our hardware matrix can help you think through.