Yes, it seems to treat each cell value as the value at a point, and create a TRN joining up all those points, rather than treating the raster as 1mx1m flat surfaces separated by vertical walls.
Does that help at all?
I think it is debatable whether this way of considering the terrain is more correct, but it certainly allows for much nicer results – try resampling your raster to a finer grid (say to 0.1m x 0.1m), without interpolation, run a 2D model based on that, and you will see much uglier results. The main downside is that I think sometimes it causes difficulties with inputting levels of things that need to be above the level of the terrain (culverts and weirs).