Physics can already be done on the GPU very well - the development we're waiting for is getting data back off the GPU and into main system memory fast enough for the CPU to be able to use it (ie, this stuff being used for gameplay, not just eye candy). That won't happen until there's a rethink on how GPUs are connected to the mainboard
This isn't the turn of the century with your new fangled AGP 4x graphics card. PCI Express is symmetric. You can pull data in from peripherals just as fast as you can push it out. If there is a bottleneck in pulling computed physics data from modern graphics cards, it's entirely the fault of the internal design of those modern graphics cards.