Reasonable GPUs

What is the status of GPUs for general compute?

Last I looked, NVidia worked well, and AMD was horrible. Right now, it looks like the major limiting factor (if you don't care about a ≈3x difference in performance, which I don't) is RAM. More is better, and good models need >10GB, while LLMs can be up to 350GB.

* Intel Arc A770 has 16GB for <$300. I have no idea about compatibility with Hugging Face, Blender, etc.

* NVidia 4060 has 16GB for <$500. 100% compatible with everything.

* Older NVidia (e.g. Pascal era) can be had with 24GB for <$300 used, without a graphics port. Not clear how CUDA compute capability lines up to what's needed for modern tools, or how well things work without a graphics port.

* Several cards may or may not work together. I'm not sure.

Is there any way to figure this stuff out, and what's reasonable / practical / easy? Something which explains CUDA compute levels, vendor compatibility, multi-card compatibility, and all that jazz. It'd be nice to have a generic enough guide to understand both pro and amateur use, e.g.:

- A770 x21, if someone got it working, could handle Facebook's OPT-175 for <$10k via Alpa. That brings it into "rich hobbyist" or "justifiable business expense" range. Not clear if that's practical.

- Kids learning AI would be much easier if it's cheaper (e.g. A770)

- "General compute" also includes things like Blender or accelerating rendering in kdenlive, etc.

- Etc.

This stuff is getting useful to a broader and broader audience, but it's confusing.