3 ms·
No, precisely because I see the emerging alternatives as being based on Vulkan and SPIR-V, which I believe have momentum far greater than OpenCL. The only quest
by raphlinus 7y ago
No, precisely because I see the emerging alternatives as being based on Vulkan and SPIR-V, which I believe have momentum far greater than OpenCL. The only question is whether a compute platform based on these standards and open source tools could be anywhere nearly as good as CUDA. I believe we're far from that today, but it is possible and there's a lot of good work in that direction.
- muro 7y agoDoesn't that exclude Apple?
- raphlinus 7y agoThat's a complex question, again the long answer is no. Both MoltenVK and the gfx-hal portability work (from Mozilla) allow you to run Vulkan workloads on Metal, thanks in large part to spirv-cross. I also see ~WebGL~ WebGPU on the horizon, and Apple is a very active participant in that process. Code written today to run on Vulkan, using any language that compiles to SPIR-V, will be pretty easy to adapt to the WebGPU future.
- ZeroCool2u 7y agoYes, though the MoltenVK project provides a Vulkan->Metal translation layer.
- ZeroCool2u 7y agoI'd really like to see the ecosystem move to Vulkan and SPIR-V as well. It's great to see that I'm not the only one! I think Vulkan has a lot of great moment behind it, especially since you can see some game shops porting fairly old games to it right now. Ubisoft's Rainbow 6 Siege is one example. Kind of bums me out Apple decided to double down on Metal and not just go with Vulkan directly, but MoltenVK is probably fine to deal with that. I would have really like to see all other platforms effectively unite behind a DirectX alternative finally.
- pjmlp 7y agoRegarding other platforms, Sony is still on LibGNM (and LibGNMX) and most Switch titles actually make use of NVN or middleware engines.
- liuliu 7y agoWhat do you think of optimizers such as XLA, TVM or Glow that try to smooth out underlying hardware differences?
- raphlinus 7y agoI don't follow this space closely (I believe XLA is related to MLIR, which I did mention), but strongly suspect that trying to run machine learning workloads portably is going to be one of the major drivers for building out tools and infrastructure for GPU compute.
- VHRanger 7y agoNote that openCL can compile to SPIR-V, so as long as Vulkan is used as the low level cross platform API we don't lose it
- gbrown 7y agoWhat about the work using SPIR—V as an intermediate representation for OpenCL? If anything, it seems like we’re moving towards a common foundation for GPU compute, with an expanding and diversifying array of higher level tooling.