>>53905479
Cuda isn't bad. I use it for doing machine learning research using different neural network frameworks that support it. As long as you know c++ you should be fine and should be able to get it after reading up on it.

>>53905479
Not CUDA but I've done programming with OpenGL Compute Shaders.

Nvidia's Compute Shader compiler seems to suck ass though, and all my attempts to use it to speed up my calculations so far have resulted in slower code.

So meh.

>>53905886
>try to speed up calculations
>get slower code
Sure nVidia like to do this in all fields, it seems.

>>53907742
It does real stupid things sometimes, like:

>Accessing a NxN 2D array is slow as fuck, but accessing a N*N 1D array (using x*N+y) is fast.

>Accessing shmem arrays with local ID dependent lookups is slow as fuck, but accessing same-sized textures with local ID dependent integer lookups is fast.

The compiler just fucking sucks. I want to try doing it with Vulkan to see if I can't write my own better optimized assembly.

>>53905479
For deep learning its pretty rad, for sheer number crunching it sucks ass compared to cheaper amd offerings.

>>53905886
>resulted in slower code
Me 2

>>53907962
It seems like the compilers's linker method is fucked up as you describes the problems