Well, yes and no, though many things like visibility checks, lighting and shaders are becoming available in the GPU, there are many things that just cant be replicated on the GPU efficiently.
Matrix trees, for example. They cannot be done efficiently on the GPU, because for almost every matrix calculation on the GPU, the CPU must give the GPU information on how it needs to be done(although matrix buffers are a viable option, just not avaliable in the forseeable future).
Physics is the prime example. Physics can benefit hugely from SIMD but not so much from SMP. The problem is that in its most accurate state(and most likely to be used in the future), events such as collisions must happen in order, otherwise you'd get freaky stuff like a train hitting someone after they had jumped out of the way(presuming this all took place in a single physics step).
Once perfect physics is achieved, im sure we'll start looking for accurate ambient lighting and raytraced effects, neither of which can run on GPU