Features
Here are the features and benefits of the AGEIA PhysX processor, taken directly from the ASUS PhysX P1 box:


These features can all be summed up by stating the four main advantages the AGEIA PhysX PPU offers.

1. Memory Bandwidth – The PhysX PPU provides for a HUGE amount of memory bandwidth, somewhere to the order of a maximum possible 2 terabits per second. This is several times more than what we see on modern CPUs. All of the data stored in the dedicated GDDR3 memory for each particle can now be collected and sent through the processor much more quickly than on a CPU. This accommodates for the massive amount of scale necessary for producing that extra level of realism.
2. Optimized Physics Processing – Those of you that have studied physics and dynamics will note that most of the calculations in the field involve the use of the use of Newtonian mechanics. Newtonian mechanics utilizes several basic formulas that are then specialized for whatever application. The calculations used to evaluate these formulas are primarily based on linear algebra. The PhysX PPU is optimized for geometric and linear algebraic calculations. This allows objects to uphold their physical shape and react in true dynamic fashion as we would see in real life. This attributes to the Fidelity involved in producing realistic effects.
3. Multi-Core Processing Complex – Details of this aspect of the PPU are not very clear, expectedly because AGEIA would not benefit from giving away their secrets. Apparently there are many individual processing elements on the PhysX PPU that can be used to compute the different physics calculations simultaneously. Almost like a bunch of small parallel cores making up an overall serial processing unit, kind of reminiscent of the architecture of a Cell processor.
4. Memory Architecture – Having an extremely high potential memory bandwidth is not enough for a processor that hopes to be most efficient at physics calculations. Objects in physics are inherently moving and changing. Likewise, data for these objects, including shape, size, and surface friction, are also constantly changing. The memory for the PhysX PPU had to allow for this dynamic data characteristic. The solution was to rework the memory architecture to accommodate for increasingly random and spread-out memory reads from the bank.

As we have stated time and time again in our other articles, theoretical features and potential performance is all very interesting stuff and makes for a good conversation, but if the realized performance of the card does not live up the expectations that that doesn’t do anybody any good.