Embed Notice

@mia@movsw.0x0.st @shibao@misskey.bubbletea.dev as much as i don't like nvidia, at least with nvidia cuda works on pretty much any nvidia card thats still in driver support, you might not get newer fancy features/optimizations on older cards, and applications may or may not work depending on how old the card is (i'd know i tried to make pytorch work on a 780ti once) but cuda itself will be there

nvidia at least doesn't do the amd thing of only supporting ROCm for the high end of the current generation and nothing else and just randomly dropping support for cards whenever they feel like it

(like, i have a 6650xt and it basically only works with rocm because functionally its similar enough to the 6000 series cards that were supported that it'll work with an env override, except it seems amd is quietly dropping ROCm support for the 6000 series in general now too (despite the 6000 series still being recent... that'd be like nvidia dropping support for cuda on the 3080 its kinda absurd)

i recently tried to make a thing work and ran into a library that supported rocm but didn't have any builds for the 6000 rx series (i didn't think supporting the currently only previous generation is that much of a stretch but okay) and i would have had to compile it myself, but then all of my attempts to compile it myself ended in suffering and failure too)


honestly congrats to AMD for having a gpu compute toolchain annoying enough to make me legitimately consider the possibility of getting an nvidia card down the line just to save myself some hair pulling as support for the 6000 series gets quietly dropped from rocm