This adds some spice to the mix...
When you have a split FE/BE situation, you have something extra - the network. We sometimes describe the idea of opening a recordset from FE to BE as a "persistent" connection. This persistent connection is usually something simple like a select query to a one-record scratch table. We only close the persistent connection when we are about to close the FE file (and thus release the BE). But I can say with absolute certainty (having tripped over this too many times) that frequently opening and closing all connections from FE to BE is a slow process.
It gets worse when your network isn't yours to control. For example, if you have an encryption requirement imposed by your local security folks, frequently opening and closing a DB requires you to renegotiate every aspect of the connection including such things as key exchanges and credential exchange - unless you happen to include a persistent connection that Access can use for subsequent connections.
Another issue is the nature of the linking between the FE and BE. I learned the hard way (from answering too many frickin' dialogs) that you get the best results if you link the BE tables but you create FE single-table queries to actually do anything in the BE. The queries remember where the tables are located and are silent when you open them.
One last "gotcha" is the depth of the path on the BE side. While I would try very hard to avoid putting a table in the top-level directory (i.e. C:\database.MDB), I will say that every layer of directories means one more set of access (little-a) arbitrations. We try to put our BE files at no more than 3rd-layer deep, as e.g.
D:\share\production\backend\thedb.mdb
No single factor I have discussed will guarantee massive performance reclamation, but failing to address each of those issues will drag down performance a little bit when considering FE/BE split databases. I'm sure my colleagues here can add further factors to the mix.