[modeller_usage] What's the difference of "refinement level" in MD

As subject, how much difference does the FAST or SLOW refinement level effect the final models in MD? Does it really worth for a loooong computation to build a more accurate model?