GPU acceleration

I am trying to run a model about hydrogen diffusion by using GPU acceleration, but the result is completely different with running on CPU in visualization. It seems that the result on CPU is as expected. I am wondering if the different result caused by precision difference? The version of LAMMPS is 29Oct20. I would be very grateful if anyone can answer my question. Thank you so much!

Impossible to say in such generality. Visualization is not a good indicator in the first place.
MD simulations are “chaotic”, i.e. even the tiniest difference (e.g. due to floating-point math being non-associative and thus the result of summing forces or energies depending on the order how the individual contributions are summed up) will grow exponentially. You have to look for indicators and properties that are independent of the starting conditions and can be time and system size averaged.

Thanks for your kind reply.
I am trying to use one pair style in my simulation to avoid the difference. The modified model is briefly described as follows: The bottom is an iron matrix with grain boundaries and dislocations which is divided into two parts, one is been set-forced and another is gave an initial velocity. The Hydrogen part is created by molecule command (Located above the z direction of the iron matrix) , and gave a initial velocity.
I am wondering if the periodic boundary conditions and NPT could simulate the diffusion of hydrogen under external pressure? Hydrogen running from one side of the box to the other can be regarded as flowing under periodic boundary? Or is there a better way to simulate the diffusion of hydrogen in iron under pressure?
Best wishes,

Sorry, but that is beyond the scope of this forum category (and my area of expertise) and has nothing to do with the topic of GPU acceleration. You are effectively asking me to tell you how to do your research. That is a discussion to have with your adviser/supervisor and possibly senior/experienced colleagues. Or you need to study the relevant published literature and compare the approaches used there with yours. Or all of that.