Is this the best path to improving computational power efficiency in the next decades?