For example, Core Image, which was part of 10.4, let developers use the GPU to perform operations [both pre-defined and new, custom operations] on images [or other data that you stuff into an image]. So the GPU manufactures are clueing into what Apple's been doing for 3 years now...

in above post.

It's all right here...
"GPUs are often restricted to linear processing, except in cases of graphics software and a handful of other programs. Apple's solution would likely involve parallel processing even in the case of routine functions"

@Guest, this technology is not new for general PC applications either, therefore no, GPU manufacturers aren't taking hints from Apple here. While Apple was among the first companies that utilised shader power for image manipulation, it didn't offer anything in the field GP-GPU, that is general computing.

One of the first GP-GPU applications was audio fx engine that was working on GeForce FX. Its creator couldn't afford dedicated pro audio hardware for his studio work, so he used untapped power of GPU. Note, the main difference between GP-GPU and GPU image manipulation is that GP-GPU guarantees correct numerical results, when just using shaders on images in general does not (precision isn't required because computational result won't be reused, it is just for screen output). CUDA requires G80-class hardware, because G80 corresponds to IEEE numerical standards (rounding, etc.). So basically CUDA is a sort of programming language that let's programmer assume he is working with general processor which is very powerful fast i heavy parallel computations.