Lets say that there are 1000 objects in a scene, all visible, which are translated by a single modelview matrix, which is NOT the identity.
Do you get anything at all out of doing the math 1 time, by changing the actual geometry, and replacing the modelview matrix with the identity?
There are no pushing-popping issues here, since it all uses the same matrix.
What I really want to know is, do graphics cards handle geometry through an identity faster than through another transform? Are they smart enough to not multiply by 1 and add 0 ?