Recently, we have found that the potential energy calculated under CPU and CUDA mode are different, which influence only the value of energy but not the dynamics of system. This bug is reported in the thread “Energy mismatching in CPU and CUDA modes for lj/gromacs/coul/gromacs”.
However unfortunately, this unaltered dynamics only holds in NVT ensemble. For calculations in NPT ensemble, everything changed, and long-term simulation shows the results are distinct from each other between those under CPU and CUDA mode.
We suppose that, the incorrect potential energy or total energy might have been used in the NPT thermostating, which therefore changed the dynamics of the system.