I believe the distinction they're drawing is they're also not "cast" to float32 during processing (which I also didn't know)