[ceptimus]

From the data you have it looks like CUDA is about two-and-a-half times slower than CUBLAS and ANN is hundreds or thousands of times slower than CUBLAS on the same number of points.

Unless there are some other factors involved, CUBLAS is the winner.

None of them looks like they will scale very well to very much larger data set sizes without taking unfeasibly long run times. Its just possible that the curves might flatten out eventually but it seems unlikely to me based on the data you have. If it were possible to do so, I'd try out just CUBLAS on some bigger data sets to see whether its curve looks like it might ever level out.