Why are folks so upset that I didn’t order some old graphics card off eBay and build my own PC to learn AI?
If you don’t like that I’m a retired web dev who can afford to spend $4k on an impulse purchase to learn AI you don’t need to stock and mock me on LinkedIn.
As if building my own PC and getting used graphics card is going to use less power?
The trolls say “what will happen when CUDA 14 comes out and NVIDIA abandons your OS”?
Fair question I guess. What is the LTS and upgrade policy on this box? So far I’ve just used the native packager manger commands and uv pip – I’m I going to be crying “why had NVIDIA forsaken me” in a year when new CUDA comes out?
I got JAX, open CL, Cuda Quantum, CUML for scikit – all just worked out of the box and see the GPU… like that is amazing, especially out of a brand new box. This has got to be easier that getting old graphics cards off ebay?
On a flip side, a lot of people seem to buy DGX Spark without knowing its limitations and then go complain when their models are running very slow. After all, DGX Spark fills a very specific niche, despite being marketed as a desktop supercomputer, and in most other cases you’d get much faster and cheaper system if you buy/build a PC with a decent NVidia GPU (doesn’t have to be used), and more stuff will “just work”.
Now, if you need to work with models that don’t fit into consumer GPUs, need CUDA and don’t want to spend big bucks or build power hungry multi-GPU systems, then Spark makes sense.
But you’ve got to know limitations - on the GPU compute side it will perform on a level of mobile RTX5070 with 70% of its memory bandwidth (or about less than 1/3 of RTX4090 memory bandwidth, about the same as Apple M4 Pro). It’s also an ARM-based system, which is generally not a big issue nowadays, but you need to be aware of it. And, being that new, not all stuff works well yet. Some of that is related to Blackwell architecture, some of that is unique to Spark.
The sparks are good training wheels for those that might build their own in the future. As someone curious and open to potentially building my own down the line I’m looking forward to learning with the Spark bundle and know what I need to build from there. I’d be curious what the trolls have in their possession first before giving any of them any audience.