![[H]ardOCP](/images/blank.gif){kind=small}
- Date:
- Monday , October 19, 2009
- Author:
- Mark Warner
- Editor:
- Brent Justice
- Google +1
Batman: Arkham Asylum & PhysX Gameplay Review
In Batman: Arkham Asylum, the Dark Knight pummels an endless stream of thugs, but will the game deliver the same two-fisted brutality to your video card? Find out as we examine gameplay performance and image quality on nine of today's video cards. Extensive testing of NVIDIA's PhysX and what it means to your gameplay experience!
PhysX in Batman: Arkham Asylum
During our gameplay analysis, we saw exactly what visual treats PhysX had to offer for Batman: AA. After that testing was complete, we have no doubts that PhysX brings real improvements to this title. But we were still left curious about some performance issues we experienced, so we explored them further.
The first thing that we noticed was that the game had some very specific recommendations regarding video card configurations when you modify the PhysX Setting on the launcher:
PhysX At "Normal"
PhysX At "High"
*NOTE: NVIDIA has advised us that the above video card recommendations are based on the game operating at 1920x1200 with no AA.
NVIDIA has told us a number of times that the best "bang for the buck" for a dedicated video card for PhysX is the GeForce 9800 GTX+, also known as G92, also known as 8800 GTS 512, also known as GTS 250. For our testing, we used the latter. But first, we wanted to make sure that the GTS 250 was really up to snuff. So, we tested it against an NVIDIA GeForce GTX 260 with 216 shaders using the same settings and the same driver: ForceWare 191.07.
So, performance is almost completely identical between the two, barring some statistical anomalies. The GTS 250 is clearly plenty of power for what we are dealing with, at least in Batman: Arkham Asylum. With that knowledge, we set about executing the rest of our testing with the GTS 250 as our dedicated PhysX GPU wherever we chose to use one.
Of course, there is a problem with the data above as you can see. At one point during testing, performance plummeted suddenly during both tests. This problem was persistent and fully reproducible. When we contacted NVIDIA about the problem, we received the following response:
I heard P35 mobos may drop to x8 PCI-E speed when two GPUs are plugged in. I wonder if that is causing it.
Do you see the problem with only one GPU? Another platform would also be helpful to crosscheck.
Lacking a second platform to test on, I went about trying to find specifically what was causing the issue. The first thing I tried was to disable AA:
As you can see, disabling AA increased framerates for a good portion of testing, but it did not alleviate the performance killing problem we were experiencing.
The next step was to lower the resolution:
So here it is. No performance hit at 1920x1200. There isn’t even a hint of the devastating blow that our performance was inflicted with earlier. At this point, we supposed that our NVIDIA contact might have been right, and that maybe the high resolution of 2560x1600 combined with a dedicated PhysX card may have been bottlenecking on our PCI-Express bus bandwidth.
However, we wanted to run one more little test just to see. So, using the "technique" learned from our community, we managed to make dedicated PhysX acceleration work with our AMD ATI Radeon HD 5870 video card. We then re-ran our test at 2560x1600 with 4X MSAA, 16X AF, and High PhysX setting running on the GeForce GTS 250.
Now isn’t this interesting! In Batman: Arkham Asylum, PhysX, which was officially disabled on AMD graphics cards, runs better with a Radeon HD 5870 graphics card than a GeForce GTX 285 when using a GeForce GTS 250 as a dedicated PhysX accelerator. Performance in general is slightly lower using the Radeon for graphics, but at least it did not suffer the debilitating performance hit.
Now that we had a good idea what we were looking at, it was time to look at some more focused metrics.
GeForce GTX 295
The GeForce GTX 295 has two GPUs, which should not come as a surprise to very many of our readers. In our conversation with NVIDIA about PhysX and SLI, we learned that incoming CUDA programs are automatically routed to the "secondary" GPU in an SLI configuration. So, the primary video card or GPU will be running only graphics, and the secondary video card or GPU will be running both graphics and CUDA programs, which is what PhysX is. In that way, SLI scaling is diminished somewhat, but with the tradeoff of PhysX processing. Since we saw nonexistent multi-GPU scaling with the GeForce GTX 295, we thought it would be interesting to see if we could extract any better performance out of it. So, we tested it with multi-GPU rendering enabled and disabled.
Unfortunately, we were not successful. Disabling multi-GPU acceleration did not enable us to set the PhysX processing to a specific GPU, so a single GPU had to do the jobs of both graphics and PhysX.
AMD’s ATI Radeon HD 5870
Here, we directly compare the efficacy of a GTS 250 for dedicated PhysX when using a Radeon HD 5870 as a graphics adapter, compared to using only a Radeon HD 5870 and running PhysX on the CPU.
As soon as we opened the door to the "outside" where the asylum was disintegrating under the gaze of the Scarecrow, performance tanked. This is the most intensive area for physics acceleration, with hundreds or thousands of bricks and papers rolling around in the wind. That is in addition to the already serious amount of GPU power used to render the many static objects, complex shaders, large draw distance, and full-screen effects.
Without a GPU running PhysX, performance is just plan terrible. But with the 3rd-party patch to enable PhysX with an ATI GPU running graphics and the GTS 250 running PhysX, everything was smooth and quite playable.
NVIDIA GeForce GTX 275
There has been some contention for a long time about whether or not running PhysX programs on the same GPU as graphics will cause some performance degradation. For this test, we ran Batman: Arkham Asylum at 2560x1600 with 4X MSAA on the NVIDIA GeForce GTX 275. We ran this test three times, one for each level of PhysX options in the game: Off, Normal, and High. PhysX was configured to run on the GTX 275 GPU in the NVIDIA control panel.
As we can clearly see, there are distinct differences in performance between these three PhysX settings. Obviously, PhysX disabled gives us the best performance. At the Normal setting, performance is almost good enough to be playable, but is just too sluggish in the early parts of our test scenario to be acceptable. Finally, PhysX at High again shows us a very sharp performance penalty, which looks very much like a bug to us. Regardless, it is rendered unplayable well before that happens anyway. So there it is: PhysX on the GPU will degrade your graphics performance as Kyle has argued for years now.