3 ms·

>> GPU may allow creating a cross-platform compute stack That was the intention of OpenCL.

collegeburner 4y ago
except writing in opencl is a dogshit experience compared to cuda. khronos is incapable of producing an API that's even remotely pleasant to use, vulkan suffers from the same thing. it is one of the worst "design by committee" groups.
- belval 4y ago
  +1 to that. Nvidia isn't winning the AI accelerator race because it's making obviously superior products to AMD/Intel. It is winning because CUDA is easy to use and reason about. OpenCL's API is terrible to use and even if you know all of it's kinks there is much more boilerplate which turns a small script into a tangled mess.
- mmis1000 4y ago
  I think vulkan is intentionally to be like that. It's by design a `not so abstract` abstraction that allow you to extract most power from the device while keep your most setup the same.(you still need to handle different vendor device specifically) It never mean to be a `write once and behaves the same on every vendor device` API. That isn't what vulkan mean to do. If you want a experience like that, opengl is the way to go.(At the cost of performance)
  - dontlaugh 4y ago
    The more ergonomic APIs with similar capabilities are Metal and it’s near-clone WebGPU.
    - mmis1000 4y ago
      I think the simplicity of metal(and the apple version webgpu, now webmetal) relies on the fact that it only has to support apple hardware (With much less edge case need to handle). I would be more curious about if the whatwg version webgpu can still be simple like that and has similar performance like the original one apple proposed. Besides that, I think DX12 is a more proper competitor about access low level primitive from devices from different vendor.
  - mort96 4y ago
    OpenGL is officially deprecated on all of Apple's platforms, and seems unofficially deprecated from Kronos; the last update to the standard was 4.6 in 2017. It will keep working (at least on most non-Apple platforms) for a long time still, but it's not going to necessarily give you access to everything your hardware can do. That's not to say it's not the "way to go", but it's not in a great state.

except writing in opencl is a dogshit experience compared to cuda. khronos is incapable of producing an API that's even remotely pleasant to use, vulkan suffers from the same thing. it is one of the worst "design by committee" groups.

belval 4y ago
+1 to that. Nvidia isn't winning the AI accelerator race because it's making obviously superior products to AMD/Intel. It is winning because CUDA is easy to use and reason about. OpenCL's API is terrible to use and even if you know all of it's kinks there is much more boilerplate which turns a small script into a tangled mess.
mmis1000 4y ago
I think vulkan is intentionally to be like that. It's by design a `not so abstract` abstraction that allow you to extract most power from the device while keep your most setup the same.(you still need to handle different vendor device specifically) It never mean to be a `write once and behaves the same on every vendor device` API. That isn't what vulkan mean to do. If you want a experience like that, opengl is the way to go.(At the cost of performance)
- dontlaugh 4y ago
  The more ergonomic APIs with similar capabilities are Metal and it’s near-clone WebGPU.
  - mmis1000 4y ago
    I think the simplicity of metal(and the apple version webgpu, now webmetal) relies on the fact that it only has to support apple hardware (With much less edge case need to handle). I would be more curious about if the whatwg version webgpu can still be simple like that and has similar performance like the original one apple proposed. Besides that, I think DX12 is a more proper competitor about access low level primitive from devices from different vendor.
- mort96 4y ago
  OpenGL is officially deprecated on all of Apple's platforms, and seems unofficially deprecated from Kronos; the last update to the standard was 4.6 in 2017. It will keep working (at least on most non-Apple platforms) for a long time still, but it's not going to necessarily give you access to everything your hardware can do. That's not to say it's not the "way to go", but it's not in a great state.