There is a lot I could say about the difficulty of programming the GPU, but I think that Folding@Home has a pretty complete thread about that. GPU programming is an application issue that would have to be tackled by the application developer. If somebody ever develops a GPU application and approaches us, we could then wrestle with the problem of boarding it.

What a nightmare that could turn into.
But interesting.
Lawrence

We really can't. While it is possible, the amount of computing power isn't as huge as advertised. Graphics is a specialised subject, with specialised hardware. Some operations are very fast, and others slow or not available at all.

All in all, the effort outweighs the benefits for the majority of projects, unless somebody creates a "Virtual GPU PC", and even that will only reduce the effort required, and won't give the massive boost you are looking for.

I still doubt that. GPU application need to be designed in such a totally different way that the porting effort will not only require a complete rewrite, it will also require a complete redesign. Then, there are more serious problems:

Most GPUs do not meet the IEEE floating point specifications. They don't need to - "almost" is close enough for a pretty picture, and accuracy is sacrificed for speed.

Many GPUs do not support 32bit operations throughout. You hear wild claims about 64bit and 128bit processors (and more), but it's not where it counts. As far as I know, no GPU can do floating point operations on 64bit numbers. Some science applications need this precision.

Taking all this into account (plus the wildly varying architectures of different chips) there's little to be gained even thinking about it at this point. This may change, graphics chips are becoming more general purpose, but it will be several generations before they are a practical proposition (and they may never be - specialisation can be good, and there is really no pressure on the graphics companies to produce a general purpose chip - quite the opposite, really).