Why are we still using CPUs instead of GPUs?

I find this a really interesting area.

As a programmer rather than a gamer I'd tended to have this background awareness of GPUs getting more and more powerful, and saw all the Amazing GPU Stunts for {SETI,Folding,..}@home or Crack Your Neighbours' WiFi Using Just A Perlin Render-Renderer, and had this vague idea that the graphics card was this generic powerhouse that just needed a little love to get going. The C-like syntax of your typical GL shader language is particularly encouraging in this regard (not to suggest for a moment that Looking Like C is a generally positive attribute).

Recently I have been giving it a little further inspection, and of course it is not quite so cut and dry. Given the Single Instruction Multiple Data (SIMD) nature, you are somewhere pretty good, roughly between a programmable DSP and Full Gory Petaflop Gigacluster.

BUT. You get:

- No recursion, but maybe we can just be careful and collapse everything into loops.

- No memory management (yay, maybe), but that makes sense since we're trying to execute everything in parallel. Perhaps we can squeeze our dynamic programming into a local 4x4 matrix?

- Expensive branching, especially particularly very expensive if it's data-dependent branching which goes different ways in different kernels. This definitely makes sense if you think of the GPU as a hungry hoarde of soldiers marching along in lockstep, slaying the data, then somebody has to stop to handle an edge case and everyone else waits around ... I'd kinda envisaged it more as an N-microthread model with a bunch of workers churning through vertices and fragments as they become available. Probably it is time to take a break from web services until I get over this, or develop an SOA GPU In The Cloud (GPUaaS?).

- &c. &c.

So, obviously there are some tasks which are well suited for this, and idioms which are well supported. Other tasks and other habits don't fare very well at all. Maybe an amazing recursive raytracer which was optimised for C is not going to survive a transition to GLSL without some very clever trickery, but perhaps there are subtasks which can be handed off instead.

The very interesting work being done on WebGL (Webkit, Firefox, Opera thankyouverymuch) seems to have the potential to bubble the GPU to the surface where it can be used and abused with a very small activation energy. I think it would be nice to have some shader skills in the toolkit just in case the right problem comes along.

TL;DR: GPU programming is a lot more like learning a new programming language than I thought.

Because it works in all cases to sell the product, you dummy.