A big problem in programming special graphics processors is that memory is very limited and can only be statically allocated. For this reason, recursive implementations are not feasible.
Because of the size restrictions, we must break large control meshes into smaller ones that the GEs can handle. Further, these small meshes have to contain all the information they are going to need so they can be processed independently. This usually means that the small meshes need to at least partially overlap.
There is also another reason for partitioning the big task into a set of smaller tasks: parallelization. In case we have several GEs running in parallel, as is the case with the Reality Engine, the submeshes should be of as uniform size as possible to avoid bottlenecks and to obtain maximum speedup.