Hmm I read this:
Peak Instruction Bandwidth: 20 billion per second
Which may indicate 20Gflops, which isnt much really, however, I can imagine it takes more power to do the same thing on a GPU as compared to a custom physics chip, like the way GPU's have to do extra calculations in Folding@Home.
So, taking that all in I would expect it to take somewhere between 30-40 Gflops for physics calculations equal to a PhysX chip. Thats maybe 4-5% of the the GTX 280 GPU's power, and about 3% of the 4870s power.
Of course, theres other factors there that will effect actual performance, like how much bandwidth its eating up, and on chip factors that could make big differences, so that really doesn't show the whole picture.