When you buy an expensive Nvidia device, you expect it to feel like part of the Nvidia ecosystem.
You expect CUDA to just be there, the full SDK in place, the latest libraries lined up and ready to cooperate.
The DGX Spark does not care about your expectations.
Nvidia still hasn’t released a complete SDK for this machine. Core pieces like Triton and downstream dependencies like JIT simply are not available yet. You also can’t load the optimized CUDA 12 stack unless everything runs inside Docker, which means hours of rebuilding libraries, sorting out confused agents, and hoping the next compile behaves.
At some point, sacrificing a goat to the AMD gods stops feeling like a joke.
There were moments this week when I wondered if I should have gone all in on a new Mac Pro or saved money with an AMD appliance. Anything with a fully grown ecosystem.
But the Spark gives something back.
120 of 128gb of the shared memory is fully usable. I can load huge models without sweating. Image and voice cloning pipelines run better than expected. Graph networks (even weird ones) feel at home here.
I also learned that the hardware strongly prefers Mixture-of-Experts models over dense architectures. LoRA weights and small fine-tunes run cleanly.
Once everything is configured, basic inference is narrowly fast enough to actually do things.
So yes, it is a stubborn little piece of hardware.
Yes, it made me question some life choices.
But I do think it will be a great little box in a few months, once the software settles.
The old ones in IT used to say, “Never buy the latest and greatest. Let them sort the bugs out for a year or two.”
This is another one of those moments where there are unexpected costs in being an early adopter.