Probabilistic Design: A Survey of Probabilistic CMOS Technology and Future Directions for Terascale IC Design

摘要

Highly scaled CMOS devices in the nanoscale regime would inevitably exhibit statistical or probabilistic behavior. Such behavior is caused by process variations, and other perturbations such as noise. Current circuit design methodologies, which depend on the existence of "deterministic" devices that behave consistently in temporal and spatial contexts do not admit considerations for probabilistic behavior. Admittedly, power or energy consumption as well as the associated heat dissipation are proving to be impediments to the continued scaling (down) of device sizes. To help overcome these challenges, we have characterized CMOS devices with probabilistic behavior (probabilistic CMOS or PCMOS devices) at several levels: from foundational principles to analytical modeling, simulation, fabrication, measurement as well as exploration of innovative approaches towards harnessing them through system-on-a-chip architectures. We have shown that such architectures can implement a wide range of probabilistic and cognitive applications. All of these architectures yield significant energy savings by trading probability with which the device operates correctly-lower the probability of correctness, the greater the energy savings. In addition to these PCMOS based innovations, we will also survey probabilistic arithmetic-a novel framework through which traditional computing units such as adders and multipliers can be deliberately designed to be erroneous, while being characterized by a well-defined probability of correctness. We demonstrate that in return for erroneous behavior, significant energy and performance gains can be realized through probabilistic arithmetic (units)-over a factor of 4.62X in the context of an FIR filter used in a H.264 video decoding-where the gains are quantified through the energy-performance product (or EPP). These gains are achieved through a systematic probabilistic design methodology enabled by a design space spanning the probability of correctness of the arithmetic units, and their associated energy savings.