JBox2D-Networking, Reproducibility and Determinism

So we are developing an online multiplayer game using JBox2D for physics. I’m also using the lockstep mechanism to keep everything in sync. It actually works really well on our tested systems, but I worry a bit about optimized floating point operations on different systems.

My research on this topic suggests that all modern CPUs use the SSE2 instruction set for floating point ops and that would mean that all calculations are processed in the exact same way on those CPUs.

But the question remains, can I be sure somehow? A (possible) different instruction set in the future could break the reproducibility and therefore my network implementation.

Using strictfp is not an option, because of the significant performance hit you get from using it.