Comment (by Zach Etienne):

@Ken Hui : If I were you I would first measure the level of roundoff error by performing a simulation with the initial data perturbed at the least significant digit(s), and monitoring how the result differs over time. GRMHD codes are particularly susceptible to rapid growth of roundoff errors, largely due to the conservatives-to-primitives solver. My guess is you might be surprised at how quickly significant digits of agreement are lost, especially when magnetic fields are nonzero.

--
Ticket URL: https://bitbucket.org/einsteintoolkit/tickets/issues/2516/different-simulation-result-with-different