Multi-core CPU in PhysX in action…
PostFX from the file script-post-fx.xml
NVIDIA has released the new WHQL version of the PhysX System Software. The PhysX System Software is included with all latest ForceWare graphics drivers but if you don’t have a GeForce graphics card, you have to install the PhysX System Software to use PhysX based applications such as PhyswX FluidMark. PhysX hardware acceleration is available for all GeForce 8-series, 9-series and 200-series GPUs with a minimum of 256MB dedicated graphics memory.
Download: PhysX System Software Version 8.09.04
The guys at hardspell.com have tested a system with a Radeon HD 3850 for the graphics rendering and a GeForce 9600 GT for the PhysX:
The functionality of CUDA and its implementation of GPU-accelerated PhysX processing has benefited many a GeForce user. Users of ATI accelerators lacking this incentive either use Ageia PhysX card or avoid it altogether. It has been verified by Hardspell that in an environment where Radeon accelerator(s) do graphics processing, a GeForce accelerator can be used standalone to process PhysX. Hardspell used a Radeon HD 3850 along with a GeForce 9600 GT on the same system with the display connected to the Radeon, though no form of multi-GPU graphics connection existed, the GeForce card partnered the Radeon well in processing physics, while the Radeon did graphics. Results of the oZone 3D FluidMark, a benchmark that includes routines to evaluate the machine’s capability in processing physics, showed a greater than 350% increase in scores, showing that the GeForce accelerator is doing its job.
|Release 177.83||Release 177.92||Difference|
|Soft Shadows Branching OFF||6054||5817||-4%|
|Soft Shadows Branching ON||10198||9542||-6.5%|
|FluidMark GeForce PhysX||9626||9579||-0.5%|
|FluidMark Software PhysX||1495||1503||0%|
Well… many websites have claimed that release 177.92 brings a boost in performance. Maybe in Direct3D (I also ran Aquamark and 3DMark2006 under WinXP and I didn’t see differences) under Vista. But in OpenGL, there is no performance gain.
|Catalyst 8.8 graphics drivers have been released few days ago, and today, or better this evening I’ve done some tests with oZone3D.Net OpenGL Benchmarks and a Radeon HD 3870 (stock clocks). I also used Lightsmark 2008 in order to have another OpenGL 2.0 application to confirm oZone3D.Net OpenGL benchmarks.|
|Catalyst 8.7||Catalyst 8.8||Diff|
|Soft Shadows Branching OFF||3384||2590||-23%|
|Soft Shadows Branching ON||4088||2993||-26%|
Global Performance Drop: 18%. OpenGL performance tumbled by around 26% in dynamic branching (soft shadows), around 23% in vertex processing (surface deformer) and around 4% in Lightsmark 2008. Only FurMark takes advantage of Catalyst 8.8 with a little 4% of performance boost. FurMark makes an intensive use of texture fetching and blending (ROPs) and maybe ATI has improved something in this part of Catalyst. But as said hereafter, Expreview has noticed a performance drop with a Radeon HD 4850 and FurMark. I don’t have such a radeon and then I can’t confirm this last result but FurMark score is somewhat weird…
Expreview on his side, has tested the Catalyst 8.8 with an ATI Radeon HD 4850 and compared to Catalyst 8.7, the new driver boosts the performance in games. If you have a Radeon HD 4850, Expreview recommends you update to Catalyst 8.8. But if you look at the results more seriously, you can notice that scores are boosted mainly for DirectX 10 titles. In other OpenGL or DirectX 9 titles, there is no or a little boost only. And with FurMark, Expreview has the same kind of performance drop I had with the HD 3870.
TweakTown in his Catalyst 8.8 review has also experienced a severe OpenGL performance drop with Radeon HD 3870 and Cinebench, the OpenGL benchmark derived from Cinema 4D. The Radeon HD 4870 has a drop too but less important.
Conclusion: for OpenGL applications that are performance focused, it’s better to not use Catalyst 8.8. In some cases, Catalyst 8.8 brings a little boost only and in other cases, a dramatic drop is noticed.
If you have some interesting results with OpenGL applications, do no hesitate to post a comment!
FiringSquad has published an article that compares PhysX performance of CPU versus PPU versus GPU. The softwares used for this test are Unreal Tournament 3, Nurien and Warmonger.
Read the complete article here: PhysX Performance Update: GPU vs. PPU vs. CPU
In PhysX FluidMark news, I put a graph that shows CPU/PPU/GPU comparison. The results show a larger difference between PPU and GPU but this is due to the kind of test: fluid simulation.
FudZilla has done some tests with PhysX FluidMark and has found that GeForce PhysX implementation is highly CPU dependant. I will try to look at this phenomenon (there is certainly something to improve with a better multithreading support). I hope the problem comes from FluidMark and not from GeForce PhysX….