PhysX relies on hardware acceleration for best performance. On low-end GPUs it is unusably slow for complex scenes as a result. On the other hand, a scene you need hardware acceleration for probably isn't going to work on the CPU anyway. PhysX probably has the biggest feature set (breakable things, soft things, tearable cloth etc.).
ODE is not hardware accelerated. It's also pretty old and not under active development any more. I assume this means it's less optimised than newer ones. ODE is extremely basic featurewise: you have solid objects, they have mass, they can be connected by joints, the end.
Bullet seems to be the system of choice for game engines that bundle physics, except for Leadwerks which bundles Newton. Make of that appeal to authority what you will. I think Newton is simpler than the others as well in terms of features, don't know.
However, if you really care about optimisation, you should optimise your scenes to prevent complexity rather than hoping your engine is capable of just swallowing it. A standard shooter game simply doesn't require a powerful physics engine: you don't need it anyway for player, NPC or most vehicle movement, or many weapons, and any of the above will handle boxes-barrels-'n'-ragdolls well enough for production use. Keep it simple and things will stay fast (most physics engines also don't bother simulate objects that are "dead", i.e. not moved in the last second or so, so the scale of the overworld won't affect performance on any of them). If things aren't fast it's most likely the scenario's fault and won't be fast in any of them.
A few hundred static objects of standard geometric shapes will be no trouble to any of those engines. Just choose the simplest one that meets your game's needs. (So you should probably avoid PhysX.) If your scenes regularly involve stuff more complicated than ~50 live geometric objects, this isn't really shooter territory any more and you should reassess your needs accordingly.
|
