Originally posted by: ViRGE
To be fair, a GPU would likely be pretty good at it too, but besides taking away resources from graphics rendering, I'm not sure the flow of a GPU is such that it would be a strong PPU.
It depends. The same or close to the same data-set that is used for physics modeling, could also be used for the graphics. Most games today only use a different (reduced) dataset for physics, and compute it on the CPU, because using the larger dataset would be impossible. But if the GPU could pipeline physics operations on the same dataset (say, using a subset of the pipelines), while also using them for graphics, that could be a win in terms of overall system performance/efficiency. (More than likely, it will continue to fuel the increasing prices of graphics cards, while at the same time making games more dependent on the GPU than on the CPU.)
If everyone remembers their PC consumer graphics-accellerator history - the first GPUs only handled the task of rasterization, while the host CPU was required to handle vertex computations. Something similar will happen with physics - today using the host CPU, tomorrow using a discrete PPU or a GPU as a PPU. It wouldn't be a bad idea to throw sound-processing into the mix as well, something like A3D 4.0 or something, with real occlusions and reflections for audio, based on a dataset similar or identical to the physics or graphics vertex datasets. (Perhaps with some sort of tagging, such that vertexes for detail meshes wouldn't get processed for physics or sound, only the structurally-interesting surfaces would.) Eventually, perhaps in another couple of years, we may start to see the next wave, true "virtual reality processors", which offload the computations of all of the essential needs of a virtual world, and the host PC simply becomes an appendix for overall gross control, and storage needs, to feed data to the VRPU. (Imagine Unreal Engine 4 running on something like that, perhaps.)
Btw, on many of today's GPUs, a number of the vertex processors aren't taxed to their full capability, so they likely have some excess computational power to spare for physics, although the future GPUs that would be more likely to be used, have shared vertex/texture computational pipelines. I kind of wonder if Cell could be used as a PPU fairly effectively, and how much more effectively a cluster of Cell procs. could be.