They composite everything at nearest higher integer scale, with correspondingly large framebuffer (just make a screenshot of entire screen and check the dimensi

They composite everything at nearest higher integer scale, with correspondingly large framebuffer (just make a screenshot of entire screen and check the dimensions). Then, during scan out, it is scaled down by output encoder. This method does not need GPU at all and also saves memory bandwidth; no need to save scaled down framebuffer and read it out again.