The primary advantage of the DGX box is that it gives you access to the nVidia ecosystem. You can develop against it almost like a mini version of the big servers you're targeting.

It's not really intended to be a great value box for running LLMs at home. Jeff Geerling talks about this in the article.

▲

cmrdporcupine 4 hours ago | parent [-]

Exactly this. I'm not sure why people keep drumming the "a Mac or Strix Halo is faster/cheaper" drum. Different market.

If I want to do hobby / amateur AI research or do stuff with fine tuning models etc, learn the tooling. I'm better off with the DG10 than AMD or Apple's systems.

The Strix Halo machines look nice. I'd like one of those too. Especially if/when they ever get around to getting it into a compelling laptop.

But I ordered the ASUS Ascent DG10 machine (since it was more easily available for me than the other versions of these) because I want to play around with fine tuning open weight models, learning tooling, etc.

That and I like the idea of having a (non-Apple) Aarch64 linux workstation at home.

Now if the courier would just get their shit together and actually deliver the thing...

	▲	mapontosevenths 4 hours ago \| parent \| next [-]
		I have this device, it's exactly as you say. This is a device for AI research and development. My buddies mac ultra beats it squarely for inference workloads, but for real tinkering it can't be beat. I've used it to fine tune 20+ models in the last couple of weeks. Neither a Mac or Strix Halo even try to compete.
	▲	lostmsu 6 minutes ago \| parent \| prev [-]
		[delayed]