What it looks like frame by frame is less important than the results of a complex interaction happening (and final reasonably correct settling point). Accuracy and precision as shown to a client can be be somewhat off but still look pretty much the same (when it gets noticeably rediculous -- like things falling in curves when they should go straight down -- then that is the point where you have to be much tighter).

There are also game methods to incrementally correct movements to make up for late data (and temporal delays over the internet) where a predicttion is made and a sudden late change makes continuing that prediction wrong, the system then slowly adjusts the paths to bring the clients representation back to match the server state.

I have no idea what PhysX has built into it (especailly for prediction/correction) and its most likely you will have to do those things yourself.

I would assume that the server would need to have the PhysX running in it because it is the arbiter of the real game state, but there are likely plenty of cosmetic effects which dont effect game play that the client would make use of PhysX as well.


Generally games requiring split second corrections all the time dont work too well over the internet because the temporal distortions cause to much noticable anomalies (too many obvious corrections which impact player interactions).