|
having open source gpu runtime, from api to metal, would be nice. but as you can see, the real meat of the business (the compiler) will probably never-ever be open sourced for internal political reasons. which is the most interesting component. it must be said the gpu crowd is very different to the cpu crowd. the cpu crowd trips over themselves to tell everyone about their ISA. the gpu crowd keep it very close to their chest. (the gpu isas are also quite batshit insane but thats another conversation.) you wind up with almost polar opposite experiences in terms of how these two groups interact with the broader industry. gpu people reeeaally don't want you prying your nose beyond the userspace APIs, in my experience. EDIT - to add though... that is kind of understandable, because the gpu crowd is under a lot more pressure to support basically everything and everyone. opengl, dxil, spirv, opencl - and the dense matrix of combinations. I often see people hate on Apple for developing their own API (Metal), but honestly? I totally get it. in retrospect they probably did the right thing. we have an epidemic of gpu programming "specs" and "standards", with no end in sight. i can't help but keep editing this comment with more of my hot takes: ofcourse nvidia owns the GPU programming space. they're the only ones offering anything resembling an actual platform! and they will continue dominating the space, because the rest of the players are too busy playing in their own ballpens. I think the only way to dislodge nvidia is to ape their silicon. a USSR needs to come along, steal their masks, and just make carbon copies. not a shim, just straight up copy nvidia 100%. like, you run `nvidia-smi` and it reports back "yep this is a real nvidia card". and then sell it for half the price. it would be totally illegal, but when youre a state actor You Can Just Do Things. |