techblogogy

1 min read

A DGX Spark is on my desk (the Gigabyte one)

128 GB of unified memory in a box the size of a paperback. First model is running. Longer write-up to follow.

It arrived. Mine is Gigabyte’s version of the DGX Spark, the AI TOP ATOM: same NVIDIA GB10 chip and the same software stack as NVIDIA’s own box, in Gigabyte’s enclosure. It’s smaller than I expected, and it’s the most memory I’ve ever had attached to a GPU: 128 GB unified between a Grace CPU and a Blackwell GPU on the same chip.

The Gigabyte AI TOP ATOM on my desk, next to the monitor, with my guitar’s headstock peeking in from the left

First evening, in order:

  1. Plugged it in, it booted into DGX OS (Ubuntu with the NVIDIA stack preinstalled).
  2. Put it on the LAN and set up SSH keys. It sits next to my monitor but runs headless.
  3. Pulled NVIDIA’s vLLM container, served a coding model, and pointed my agent CLI at it.

It answered. The whole thing, box to first token, took one evening, and most of that was the model download.

The question I actually care about is whether this can replace my ChatGPT subscription, and what “replace” even means. I’ll use it for real work for a few weeks and write that up properly.