In general, you should write what will be the cleanest code, and only optimize it after verifying through measurement that 1) there is a performance problem and 2) this is actually faster.
It seems like a reasonable first assumption to expect that the database does pretty much what you would do by hand, but without the double roundtrip. I think the amount of data sent would be equal when using prepared statements too.
This is not to say that the above paragraph is the answer, just to show that conflicting theories can seem reasonable and that you always need to measure.