Over the years working with many CAD tools from Pro-E/Creo, Solidworks, NX and even a little Catia it's always been a discussion on model rigging with common origins/coordinate systems. One of the draw backs that was documented especially by PTC for Pro-E was that as model extent (overall diagonal model size) versus the size of the features (very small features) contained within that model the model memory (RAM) required to load increased due to model accuracy having to increase as well. So if your model is using a common origin your model extent would increase based on far a part is from the origin.
I'm curious is there any documentation anyone has seen of this for NX (or even Solidworks if anyone has that info too). I've run test with models and found that the above holds suit in NX along with Creo and Solidworks, but find no documentation on it except by PTC for Creo.