A) The first point (the link is down atm. I I am guing to guess here) was debunked. AGEIA's PhysX wasn't either multithreaded (Unless the devs specificly coded it to be multithreaded)
or used SSE2. It was single threaded and used x87. NVIDIA added (auto) multithreading and SSE2 to the API, thus make that point: meh.
B) Comparing 3000 rigid bodies ot 100.000's of rigid bodies or comparing games where destroyed 100's of items magically dissapear after 10 seconds or become "dead" to 1.000's of rigid bodies that reamain and stay interactive is a bad comparions.
You could ask JFAMD what runs faster (or give an educated guess)...AMD CPU Bullet Physics...or AMD GPU OpenCL Bullet Physics?
Physics API's have (in theory) the same features...it's just that the FLOPS from a GPU is so much more powerfull for physics calcualtion than a CPU.
Ask yourself this:
Why does AMD want to add a "GPU" to their CPU core?