Ask HN: Optimal Setup for Home ML

Stable diffusion is awesome. I've been inspired to start hacking on ML stuffs and I'm having tons of fun just using it for making images. My m1 is ooookay at it - but it's time to upgrade.

I could drop a bunch on a 3090 and call it a day, but I really don't need to be gaming and anyhow looking at passmark(1) it seems like I could get more throughput / $ with an array of cheaper cards and a box that'll take more cards...

..Which leads to some questions, because I'm kind of ignorant.

- What metric should I be looking at for running ML pipelines? Performance(1)? GPU Compute(2)? Should I just go for the most memory / dollar?

- Are there other considerations (tensor cores etc) that go into this? It looks like the GTX series doesn't have em, so I'm assuming those are a no-go.

- Is there a reason to prefer Radeon or Geforce over the other?

- Do all the cards have to be the same, or can I just get whatever is handy at the moment and put em all in together?

- How much do I care about the CPU? This box will be doing nothing but running ML pipes.

- Finally, would I be better off saving up for an A30 or some such (in terms of power / price)?

Reading up on (3), I gather that I'll need the cards to be single-slot with SLI or Crossfire, and a big beefy power supply, along with an ATX board with four slots.

The plan is to have this thing running as a server in a closet, so I can send it jobs from the laptop. No reason IMO to bother with a monitor or other such if I don't have to.

1 - https://www.videocardbenchmark.net/gpu_value.html#

2 - https://www.videocardbenchmark.net/directCompute.html

3 - https://www.newegg.com/insider/how-to-choose-a-motherboard/