> Unlike CUDA, there's no capable API that works on the whole range
Microsoft's D3D is very capable, and vendor-agnostic, same code works on all GPUs because the user-mode half of the GPU driver includes a JIT compiler for these shaders.
Not sure why it's not popular outside of game development. IMO their GPGPU API is the best one currently available.
D3D is the exception to basically everything in graphics-land, avoiding the pitfalls that basically every other cross-hardware-vendor API suffers. The problem, obviously, is that it is Windows-only. It's only popular in cases where the rest of the system is also Windows-only, since otherwise you'd have to write multiple backends anyways (at which point you either stick to OpenCL, or you do both CUDA and OpenCL for better performance).
Being limited to Windows makes DirectCompute a non-starter to basically anything in the HPC/server/embedded space.
Microsoft's D3D is very capable, and vendor-agnostic, same code works on all GPUs because the user-mode half of the GPU driver includes a JIT compiler for these shaders.
Not sure why it's not popular outside of game development. IMO their GPGPU API is the best one currently available.