One of the benefits of using many subcomponents vs. one big component is it allows you to setup a hierarchical linear solver approach for the adjoint (to be more precise, the UDE linear system). Depending on the model's coupling structure, a hierarchical linear solver can be much faster than a "monolithic" linear solver applied to a model with a few big components. The appropriate linear solver setup is highly problem-dependent, though. This benefit is about the linear solves for the total derivative computations, and not about computing partials. So it might not be relevant to your questions on AD for partials.

Note: I'm not an OpenMDAO developer and I could be wrong.