The way I look at it, almost everything, including physics and AI, should be running on the gpu itself. If an entire game program could be run within the confines of a gpu, I think we might see a fundamental shift in the depth and immmersion possible. I'm thiniknig holo-deck style though...probably not what most expect.
We aren't far away from rael holodecks being possible...it's actually quite scary how close it is.
To do this, gpus must become a bit more compartmentalized ...as they were before... with rendering pipelines, although this way of doing things was killed almost 5 years ago now. With seperate "cores" doing different tasks, but each working together to render a final frame, real physics simulations will be truly possible.
I kind of envision, for physics, a complete rendering cycle that never stops during gameplay...and as you play the game, the game's world's physics will be running in simulation on the gpu, creating your triangles.
DirectX 11 kinda works towards this, but as many have seen themselves, DX11 comes with quite a severe performance impact...and I bleeive part of this is because gpus are really ready just yet.
You could look at a single SIMD as a pipline, but to me, they are not fully featured enough yet. I think we need quad gpus...like how tri-sli with a physics cards runs...but I want that physics card to take a far more primary role than it does currently.