[dagw]

GPU pricing is kind of all over the place at the moment, but, assuming you are on a budget, I would look for either a RTX 2070 (super if you can) or GTX 1080 Ti as great 'starter' cards for ML. Both should be available second hand for a fairly reasonable price. If you want a 'new' card, I would go for the 3070 or 3060. The 3060 is interesting since while it has few and slower cores than the 3070, it has more VRAM which is important for certain workload.

As for the rest of the components, don't sweat it too much. As to which is actually more important, it depends a lot on exactly what you are doing. 32 GB of RAM, basically any 'mid range' CPU from the past few years and an SSD an you're good to go.

One thing worth considering however is getting a latest gen CPU/motherboard so that you can upgrade the CPU in the future. For example, while you will probably find great deals on AMD AM4 CPUs and Motherboards today, you won't be able to upgrade to the latest AMD CPU in a couple of years since AM4 is end of life. Getting an AM5 CPU/motherboard will cost more up front but you will have the option to upgrade the CPU down the line.

Also as a final point, ask yourself do you really need an ML machine? If you're just 'playing around' consider sites like Colab and others that offer free or cheap GPUs for training ML models. I know lots of people who have done real ML research and publish actual ML papers using only the free tier Colab, so you can get quite far on it.

[rjh29]

You have to pay more for the CPU, mobo and DDR5 ram, so it's quite a tax right now. It depends on your workload of course, but if it's mostly GPU-bound (ML, gaming) I think the AM4 5800X3D is a really good choice that is destroying benchmarks and should last 4-5 years.

For GPU it depends on your vram needs, the Stable Diffusion stuff seems fine with 8, but Textual Inversion (adding yourself to the Stable Diffusion model, etc.) and text/transformers work needs more. I wanted the latter so I bought a used 3090 which was relatively cheap (compared to before, anyway!)

[dagw]

so it's quite a tax right now.

True, I guess it really depends on your workload. If you're not CPU bound and don't really expect to be, then getting the best CPU/mobo combo you can afford today and not worrying about upgrading might be the best way to go.

[tstrimple]

With Intel's latest release, I'm not sure AM5 is worth investing in at this point. Raptor Lake seems to crush AM5 with support from existing sockets and DDR4 all while being cheaper. That massively cooled my desire to go all in on a big AM5 system.