I get that, mostly I use them for dsp for software defined radios. The drivers for those are probably hard to compile as well. Depedencies on libusb often. Or dpdk for the nice ones.

I'm still holding out hope that Intel will put out a decent dGPU with a strong ML story within a few gens, with gloriously open source drivers.

Recent news coming out of Intel the last week or so haven't been encouraging though. But I will remain a dreamer.

AMD is probably a better bet with ROCm although it's still pretty spotty right now in regards to support.

Btw, dunno much about software defined radios and such but I'm curious. Isn't this the kind of niche FPGAs are perfect for? Or do they not justify their cost in your use case?

Purchase price yes, development costs less so. Maybe I'm just bad at verilog but it takes me a lot longer to do the same thing compared to cuda or c++. Also the devtools are a big pain, and nothing is performance portable across device brands. Or even device classes. But yes, for high rate simple-ish dsp, fpga is awesome for performance and performance per watt.