Unfortunately due to time constraints we werenít able to run SLI PhysX (where two cards are running together in SLI graphics mode with one card handling a mixture of graphics and PhysX) verus multi-GPU PhysX benchmarks (where one card is dedicated solely to graphics, and a second card dedicated solely to handling PhysX), but we did manage to get some good multi-GPU PhysX benchmarks, as well as determine a possible baseline for the minimum hardware youíll want for dedicated GPU-based PhysX processing. We hooked up a 9800 GTX+ handling graphics with 9800 GTX+, 8800 GT and 9600 GT cards that were dedicated just for PhysX. Here are the results:
As you can see, all three cards yield the same performance numbers with the 9800 GTX+, suggesting you donít need anything better than a 9600 GT for handling dedicated PhysX processing. Letís see what happens when we pair this 9600 GT with cards ranging from the 8800 GT up to the GTX 280:
With the GeForce 9600 GT dedicated solely for PhysX processing, the primary GPU is able to focus exclusively on rendering graphics. We saw a nice performance increase as a result, with the GeForce GTX 200 and 9800 GTX+ cards seeing a performance improvement of 17-18%, while the 8800 GTís performance improves by 25%. Not bad at all!
We should note that we took another stab at running the 9600 GT in the secondary PCI Express graphics for PhysX with ATIís Radeon HD 4870 handling primary graphics duties. Unfortunately however the NVIDIA graphics driver wouldnít let us do this, so mix match PhysX appears to be limited to GeForce cards only (i.e. mixing a 9800 GT with a 9600 GT).
UPDATE 2PM: We just heard back from NVIDIA regarding mixing and matching ATI and NVIDIA hardware. Unfortunately this isn't supported by Vista. The Vista driver model only supports one graphics driver at a time.