New hardware?

[EvilDragon]

The thing is that CUDA cores really do just relatively simple operations. They are nowhere near the scope of math operations that regular and specialized CPU registers can do. And while MMX, SSE and others were introduced to add to 3D rendering performance, that's not the only thing they were good at doing. SSE and AltiVec do a lot for FFT processing as well.

And ultimately, this is a major difference: GPUs need to calculate a lot of pixels for a couple dozen frames per second. That's a couple dozen. Audio operations need at least 44100 times per second. Or more, in case of oversampling. This is where that "returning values consistently" problem is occuring.