I run 2x 4890's and a 100 series PCI-E Physx PPU and i can tell you with the latest drivers it works. The fan on the PPU only turns on when Physx is being processed. Load MOH airbourne and the fan spins and maybe you might notice some minute differences when playing
I run the PCI version from ASUS. The PCI and PCI-E versions are identical - the PCI-E just has a PCI bus to PCI-E bus adapter (which is why there is no frame-rate improvement between the two models - the PCI bus creates a frequency bottleneck, even though the data-rate of the PCI bus is adequate for the amount of data being sent back and forth). The fan spins period.
That said - I've seen it confirmed on several other sites that the PPU is disabled. My own PPU cannot be located and used for hardware rendering utilizing any physx code - though it shows up in the device manager and in the Nvidia control panel.
The programs will run through and search for CUDA-enabled devices, and then the PPU. Since I am running XP64 and a 7950 GPU - my PhysX hardware mode is disabled due to a weird way Nvidia enacted their coding (cool how my Nvidia card doesn't count anymore).
It works like this: every time they update their drivers, they have to verify that it doesn't break on old hardware. It's possible that they didn't just add lines of code to disable the PPU with ATI hardware, but that they were able to remove LOTS of code that included workarounds for the old hardware. This would also cut lots of time out of validation.
This is incorrect. PhysX now runs off of CUDA - which runs completely independent of graphics processes. That, or it runs off of AGEIA's card - the specifics of which are not known. Nvidia runs into more bugs trying to get PhysX and Graphics to run on the same GPU than it would to run graphics on an ATI card and PhysX on an Nvidia card.
Bear in mind that NVIDIA has already discontinued support for the PPU at the SDK level. SDK version 2.8.1 was the last version to support the PPU. Projects built against SDK 2.8.3 or later can only be accelerated by a GPU, largely as a product of features added to the SDK that the PPU can't handle.
The reason Nvidia discontinued support for the PPU (which is actually a far superior design when it comes to PhysX processing) is because it is an old product that requires heaps of additional coding to support as it is not just an expansion of CUDA. It's more work to support the PPU and its architecture - and it's an older product that is not really good for anything outside of being an aid to simulations.
And Nvidia isn't going to make a dedicated PPU based on the architecture, even though it is a superior PhysX architecture (yes, GPUs get a higher score.... they also run at about four times the clock speed and have about eight times the number of 'shaders' - I'm talking about architecture, which is different from benchmark scoring). A new PhysX card is not going to draw Nvidia any profits any time soon, and it would simply split their development down the middle.
All this boils down to is Nvidia trying to push CUDA.
The whole reason ATI didn't take up PhysX was because for PhysX to work on ATI cards, it would have to adopt CUDA. ATI already had its competing stream-computing solution.
Of course, now that we have OpenCL, AMD/ATI have ported their format (which is, more or less, free) and started backing Bullet - it will more or less choke out proprietary formats like PhysX.
Now all I want to know is if my darn PPU will be able to utilize OpenCL. I mostly want it to help in simulations - not gaming, but I see no reason why it couldn't support OpenCL, other than it is an older product that had a small consumer base to begin with.