Well the other possibility is that the power benchmarks are accurate: the buffer did save power, but then they figured out an even better optimization on the mi

Well the other possibility is that the power benchmarks are accurate: the buffer did save power, but then they figured out an even better optimization on the microcodes level that would make the regular path save even more power, so the buffer actually became a power hog.