Well, the last issue of the c't magazine (or the one before that) had an article that discussed the computation of game physics a) on multi-processor machines, b) on machines with more than one graphics processor where the second is (mis-)used for the physics, and c) dedicated physics accelerators as the one by Ageia.
In my opinion, using a graphics processor for anything else than graphics is just toying around, you also don't use a cellular phone to listen to your latest CDs/mp3s either (even though you probably
could use it as a replacement for a HiFi station; the phone probably even has more CPU power these days).
It's nice how imaginative people are when it comes to using GPUs for general-purpose processing, but I really think that these applications are highly limited due to the specialized nature of the GPUs.
The best solution in my opinion is using multiple processors (of the generic CPU type) for this purpose, because they are universally programmable etc.
Dedicated PPUs (Physics Processing Units) may or may not have their purpose - time will tell. If I could choose, I would prefer everyone gets a multi-CPU computer rather than a PPU card,
but for me, the same if true for GPU cards. (The ideal computer needs no accelerators for anything...
)
To summarize, I think that physics in Ca3DE should be computed on the CPU. If we have the option to have it
also run on a PPU, being able to fall-back on the CPU if none is present, I'd prefer that even more.
Therefore, I only suggested to use Ageia because I believe that I've read somewhere that they run their physics on the CPU if no PPU is present (it's probably easy to verify that on their website), so using the Ageia SDK
might be a good choice, pending further and more detailed investigations.