A fair comparison between Bullet and PhysX

[guernika]

Dear All,

I developed a simple test to compare these two physics engine.
You can watch the results following this link:
http://youtu.be/plwS86ddbf4
Trying to summarize, bullet is faster, but Nvidia PhysX collision detection works better.
Maybe mesh cooking in PhysX performs some convex decompositions?
I would like to improve and extend this test, so I wait for suggestions and comments by you all.

Best Regards,

Francesco

[Dirk Gregorius]

I think this test has no value at all! You are testing just one cloth specific feature of both engines and you don't provide any information which settings you used in both engines. Nor do you describe the test configuration.

[guernika]

I just wanted to compare collision detection against triangle meshes, which is the most expensive task in terms of computational times. I don't care about cloth simulation. I think that cloth simulation solvers are quite similar, but I'm not sure of this.
On the channel there are info about triangles and vertices of both meshes, time step, and collision margins. What else do you want to know?
Finally, as I told in my post, I am waiting for suggestions to improve the test and discover some meaningful results, and your post does not help me so much.

I think this test has no value at all! You are testing just one cloth specific feature of both engines and you don't provide any information which settings you used in both engines. Nor do you describe the test configuration.

[Dirk Gregorius]

Well first of all I would you prefer you to be more explicit of what you are actually testing. So the topic of the post should have been "cloth vs triangle mesh collision comparison". Your topic suggested (at least to me) a full PhysX vs Bullet comparison. Also if the intention of your test is to test collision performance where is the result? In the video I only see the simulation time. Maybe you want to split simulation and collision performance. Finally you can add some error metrics (e.g. penetration and maximum stretching) and assure those metrics are satisfied (e.g. no penetrations) over the simulation time. Finally, PhysX also solves self-collisions. Is this enabled? Does Bullet implement this as well? I would at least expect to read which solver configuration you used for each engine. Finally, it is nearly always a good idea to open-source those comparisons, so people can make their own conclusions.

HTH,
-Dirk