I had thought this was also to do with gpu physics at some level of precision in the float it matters so one machine configuration will not precisely match another machine doing the same calculation. Or something like that.