An Early Stage Exploration of Stochastic Computer Systems

University Of Illinois At Urbana-Champaign, Urbana IL

Investigators

Abstract

Traditionally, computing systems have been designed to always compute correctly - i.e., they are designed for error avoidance under all system conditions; fault tolerance strategies are implemented primarily to ensure continued correct execution in face of unforeseen and infrequent faults in the system. This conservative, worst-case design philosophy often leads to very inefficient designs, especially in terms of power. This project looks at the feasibility of developing a new class of computing systems, referred to as stochastic computer systems that perform best-effort fault avoidance in face of other objectives like latency and power. The primary benefit of a stochastic computer system will be allowing the use of very low-power, but unreliable technologies, as well as allowing aggressive low-power optimizations.