I don't actively use it for my workloads (at least, not consciously), but my understanding is that the cache keeps stuff that would otherwise involve polling the VRAM, so that's the fast part. I tend to do most of my work with this card on Hackintosh, so there isn't the granularity of control you'd get from a proper vendor driver / suite…so, can't really comment on real world for Windows.
Stream Processors are sorta similar but different to CUDAs, you can't do a straight compare of the number of each and get the 'better' card. CUDAs are more optimised since Nvidia have wider reach and more experience, and it follows that devs will optimise for the most widely-available technology. Compute Units are collections of processing elements / components.
It's subjective, but I only hear the fan on the card when I run it red in DaVinci Resolve, and then it's just part of the overall white noise. My case (Thermaltake V1) isn't designed to be quiet, it's full of holes and is pretty utilitarian, but I have no complaints about noise.
Modular. I have 6x SATA SSDs running in a small-ish case, along with SATA power splitters and such. Again, because it's a pretty basic case and 'good enough', the airflow isn't a concern for me. There's a 6-pin power input on the top of the card, so check your headroom between the cable curve and the roof of your case. If it's snug, maybe a low profile or angled adaptor will keep it low.
UK. Second hand from eBay…well, barely used second hand, and it was in an eGPU enclosure which I flipped, so the card itself was around 375GBP in mid-2022 after I'd made some money back on the enclosure.
