I can tell you years back on NVidia there was a pretty serious perf penalty whenever you switched the size of texture that was bound as the FBO as a render target, confirmed by an NVidia guy. He suggested creating one FBO per size and I got a nice speed-up in response.
That’s was 4-5 generations of cards ago though. As the FBO abstraction layer I wrote uses separate GL FBOs per res and hides this odd quirk, I’ve never bothered to retest this on recent cards. Next time I rip into FBO perf I probably will though.
Along with different resolutions, test different formats too, if you care.