Industry will soon manufacture transistors whose sizes are on the order of atoms. These tiny transistors are fragile, meaning each transistor will be different and will change during use. Many transistors will be unusable when first built; still others will degrade or fail with use. Like people or snowflakes, each of our components (e.g. microprocessors, graphic chips, memories) and systems (e.g. cell phones, mp3 players, anti-lock brakes) built with these tiny transistors will be unique. Traditional, one-size-fits-all approaches assign tasks to transistors oblivious of their unique strengths and weaknesses; these approaches waste much of the potential benefits of these tiny transistors, leading to systems that cost too much, use too much energy, and fail too soon. This research explores a novel assignment approach that assigns tasks adaptively based on measured transistor characteristics. The fastest transistors are assigned where they most accelerate performance, while the slower transistors can still be used for less time-critical tasks. Assignments are further re-evaluated during system operation, allowing fresh transistors to replace transistors that wear out.
Practically, this means IC manufacturers can produce smaller transistors and continue to deliver more capable electronics (e.g. digital video recorders, cell phones, laptops, supercomputers) for fixed dollar budgets. These capabilities continue to improve our quality of life, providing richer media, better communication, greater automation, and greater safety. This work will reduce the energy per computational task thereby extending battery life, reducing energy bills, and facilitating cooler operation. Replacement and reassignment mean electronic components will last longer and degrade gracefully.
