Thursday Keynote: Survival of The Fittest: Circuits and Architectures for Computation with Wide Power- Performance Adaptation Beyond Voltage Scaling

Abstract

Wide power-performance adaptation is becoming crucial in always-on nearly real-time and energy-autonomous integrated systems that are subject to wide variability in the power availability and the performance target. Adaptation is indeed a prerequisite to assure continuous operation in spite of the widely fluctuating energy/power source (e.g., energy harvester), and to grant swift response upon the occurrence of events of interest (e.g., on-chip data analytics), while maintaining extremely low consumption in the common case. These requirements have led to the strong demand of a new breed of integrated systems having an extremely wide performance-power scalability and adaptation, beyond conventional voltage scaling or adaptive parallelism. In this context, systems being able to adapt to a wider performance-power range (“the fittest”) allow true continuous operation and adjustment to the power-performance profile required by the application (“survival”). In this talk, new techniques that drastically extend the performance-power scalability of digital circuits and architectures are presented. Silicon demonstrations of better-than-voltage-scaling adaptation to the workload are illustrated for both the data path (i.e., microarchitecture) and the clock path. Adaptation to a very wide range of energy/power availability is also discussed, presenting demonstrations of always-on systems (e.g., microcontrollers, power management units) with power down to sub-nW, and duty-cycled operation down to pW range. As an orthogonal design dimension, “just-enough” adaptation to the application-level quality requirement is shown to further extend the performance-power range by an order of magnitude or more. Under this energy-quality scaling framework, quality is treated as an explicit knob, eliminating the quality slack that is traditionally imposed by worst-case design across different applications (e.g., machine learning), contexts, datasets, and the pessimistic design margin to counteract process/voltage/temperature variations. Several silicon demonstrations are illustrated to quantify the benefits offered by wide powerperformance adaptation, and identify opportunities and challenges for the decade ahead.