That is generally a sub-sampling or supersampling of the DTM data. I.e. a input DTM of 20m is automatically saved at different resolutions on creation to allow areas further from the camera to be loaded at lower resolution (i.e. 2km GSD) and areas closer to the camera to be loaded at full resolution (i.e. 20m GSD) as well as provide the ability to restrict loading of higher resolutions on slower hardware.
If you consider a DTM to be like a picture where each pixel is a value which represents height, then a picture with more pixels covering the same area can contain more information about the terrain it is representing. However, since computer hardware is limited in its ability to load and render data, you need to provide a way for the user to restrict what the computer loads so that it's resources can be shared among other things (i.e. rendering the aircraft, weather, AI etc). Therefore, on a less powerful computer, setting the terrain resolution to say 40m will only load data up to half the resolution of the source 20m data.
In terms of the limits of resolution, I'm not sure about X-plane as I haven't developed for that but for FSX/Prepar3D you can store sub-meter terrain data (technically it's a value in degrees rather than a linear meter)- You'd only really see that detail when up close but I've released 5m lidar derived DTMS for FSX which work well. Prepar3D V2 is on the horizon which moves terrain onto the graphics card so I'm not sure if the resolution of the source DTM files has increased as well.