Toward Physically-Adaptive Computing

As semiconductor technology approaches the atomic scale, electronic systems are increasingly burdened by physical variations and uncertainty. Traditionally-designed systems lack an ability to adapt to these fine-grained effects and are thus becoming more inefficient, error-prone, and subject to early wear out. This paper describes the paradigm of physically-adaptive computing (PAC), in which systems learn physical parameters and adapt with fine granularity in the field. We outline an architecture for an adaptation agent and investigate two key aspects of the adaptive process: self-characterization and physical self-optimization. A case study is presented involving random variations in latch reliability. We conducted experiments on a model of a PAC system with physical data obtained from actual field-programmable gate array (FPGA) hardware. Our results show that across 15 benchmark circuits the mean time between failures improved by an average of 30% via low-cost self-adaptation and by 45% assuming assistance from a remote server. Physical self-adaptation and assisted adaptation will both play an important role in achieving computational systems with atomic-scale features.