This is a good question.
With the current implementation of the storage engine in MongoDB, the documents get entirely loaded in memory. Then the BSON structure is parsed in order to extract the different fields. This operation of parsing the document is done in memory and is really fast compared to the other operations like finding documents and index keys in the trees, which are also fast, however, these will require more computation cycles.
In other words, having an embedded structure or not is not going to make a difference for performance.
I recommend grouping information in sub-documents because it will make it easier to manage, especially on the application side where this grouping may match the objects in the application.