The gist is that the code used for PhysX to run on CPUs is mostly X87 and single threaded while it could have been SSE instructions and multithreaded. Currently when run without a Nvida GPU and having the PhysX done on the CPU instead PhysX is very slow. But it COULD be very fast...possibly faster than when run on a Nvidia card (vs a CPU w/4 cores free and current double-precision SSE compared to older CPUs). And support for the old Ageia PhysX cards is no more with current drivers so compatibility can't be holding Nvidia back. So did Nvidia pick the slowest code for CPUs to make their cards look like they run PhysX better for a selling point or do they just not care?