Comment by 15155

I don't mind CUDA, I do mind that all of the SDKs don't dynamically load the various CUDA shared libraries at runtime.. intertwining itself into your application linking process makes for extreme binary portability inconvenience.