Got into some heated debates with people about physx where from a developer standpoint it was off-loading all calculations to the CPU with the present of a physx GPU. Typically in game engines you need to pause the world and look into the physx simulation to get the object's pos/vel/mass. This causes the physx engine to stop the world to finish it's calculations so crossing the PCI bus becomes more and more expensive.
Vast majority of the time, most game engine's to solve this problem just let non interactive objects (cloud's, waves, lighting) to be calculated by the phyx sub-system, and only poking it now and then to prevent the phyx hardware from swamping to the CPU.