> running the physics engine at a deterministic timestep

As well as using special library versions of floating-point functions which don't behave the same across different processors I suppose, if you want to be safe.

Eg cr-libm[1] or more modern alternatives like core-math[2].

[1]: https://github.com/SixTrack/crlibm

[2]: https://core-math.gitlabpages.inria.fr/