Arkham City PhysX Multi GPU Comparison

So I see a lot of people saying no games use PhysX and that it's actually something of a novelty that never really took off just like EAX by Creative Labs not too long ago.  Seeing how phenomenally the recently released Borderlands 2 has made use of it I have to disagree with sceptics.  It isn't perfect, but it definitely makes for a much prettier and more realistic environment.

Above you see my set up of two GTX670 SLI with a GT430 and GTX480.  Unfortunately I no longer have access to the latter 2 cards used in this test, therefore I can't run a test measuring Borderlands 2 performance.  I do however have results testing Arkham City.  I've chosen to test under DX11 with everything else turned to maximum setting (obviously PhysX cranked all the way up as well) vsync off at 2560 x 1440 with 4xAA because it is the most intensive scenario within the limits of my system, and more importantly, its what I'd play the game using.

Click to enlarge.

The results were quite interesting.  The minimum fps was quite sporadic across the various combinations, but were consistent through multiple runs of each.  At first I thought it may have something to do with the two cards connected using an SLI bridge.  But after using multiple bridges and a test of dual GTX670s with one dedicated to PhysX and without a bridge, there were no differences in the minimum fps.  What also threw me off was the fact that a single GTX670 + GT430 for PhysX had just as low a minimum.  These must be anomalies though as there was never a point where I'd see it running at 1 fps for any sustained period of time.

Anyhow, apart from that little anomaly, the average frames and maximum frames give a much clearer indication of PhysX performance.  Running with a single GTX670 and offloading the PhysX to the CPU was surprisingly playable (Core i7 3820 @ 4.6GHz).  It would drop below 30 more often than I would deem acceptable but it definitely wouldn't come to a sustained crawl at any point.  But moving on to the more realistic usage scenarios and we find that a single GTX670 does a fair bit better on average than CPU PhysX.  Adding a dedicated PhysX card offered extremely incremental advantages over a single GTX670 with a GT430 actually lowering performance by a couple fps on average while the GTX480 showed 1-2fps increase over the lone card.

Moving onto SLI scores and we see a similar pattern.  Whilst the averages remained fairly lacklustre the maximum did  jump about 20fps.  Adding the GT430 to the SLI GTX670's again saw a lowered average score whilst showing no real world benefit.  The GTX480 on the other hand helped a hell of a lot more in SLI pushing the average above the 60fps threshold and the maximum above 100fps.

In conclusion, those opting for a dedicated PhysX card, especially the AMD users, don't bother with any card that would run significantly lower than your main GPU.  The GT430 did no good for PhysX at all and it'd be better saving that money and going single GPU.  For SLI users, a strong dedicated PhysX card does help, but the extra heat, noise and power draw from having a GTX4/5/670-80 solely for that purpose, not to mention the extra wad of cash you'd be without, is definitely relative to what you value most.

My pick?  No dedicated card at all.  But if you've got SLI and can afford it:

One of these might help =)