Quote Originally Posted by trinibwoy View Post
What would be the point running a serial program on a GPU? The problem isn't cuda or opencl. The problem is that either the math isn't parallelizable or people aren't smart enough (yet) to make it so.
This is correct.

The math IS parallelizable. And the proof is here (already posted):

http://www.xtremesystems.org/forums/...d.php?t=221773

HOWEVER,

It is very difficult to do. It was hard enough to parallel it into several threads for several cores. So it will be MUCH harder to parallel it into hundreds of threads for a GPU...

Also GPUs right now don't have the right instruction set to be efficient with this kind of computation. (They are too specific for graphics.) So I won't be surprised if even the best of GPU implementations have trouble beating a CPU implementation.