>>42
Sorry, yes, nVidia supports OpenCL on their older cards too, but I meant the divide between CUDA cards and OpenCL-only cards. But it doesn't matter--I am not interested in portability even a little.
As far as the compiler goes, it doesn't matter what it is, it matters what the code it produces actually does. For example you could compile C to Java bytecode plus some runtime environment, but the system you're targeting is heavily abstracted. I don't think OpenCL is that bad, but if I recall correctly it is significantly more abstracted than the CUDA scenario. You don't seem to know so I'll stop asking and trust the conclusion from my memory.