I also had a slow down when tested a few hundred vertex animated soldiers with runtime attached randomized heads, weapons and shields. it really does depend on how you calculate relative positions. cloning also can result in slow down.
if use only vec_for_vertex and get the same pan, it is relatively fast. if more complex vector calculations are applied it can be felt...
if you have bone animations, you can use vec_for_bone and ang_for_bone (but I have little experience with it).
if entities are animated you can merge them in MED and use vmask runtime to show only what you need, or use wmb if they are non-animated (but merging them in MED also fine).
but it really depends on your particular application, which is the best way...