Probabilistic Value-Deviation-Bounded Integer Codes for Approximate Communication

M computing systems are designed to prevent errors in computation, memory, and communication. Guarding against errors however requires energy, temporal redundancy, or spatial redundancy and therefore consumes resources. But not all systems need to be free of errors: In some systems, either by explicit design or by the nature of the problems they solve, system output quality degrades gracefully in the presence of errors. Several important future applications of computing systems, ranging from wearable health-monitoring systems to neuromorphic computing architectures (1) often dissipate a signi cant fraction of their energy communicating data (e.g., from sensors) in which the e ects of errors are best quanti ed in terms of their integer distance, rather than using a Hamming distance. At the same time, the computations that consume this data can often incur errors with a wide range of integer distances with limited system-level consequences. Because electrical communication interfaces do not bene t as much as processors and other digital logic from semiconductor process technology scaling, the fraction of system energy dissipated by sensor data access and communication is currently large (2–4) and will only grow in future embedded sensor-driven and neuromorphic systems. It is therefore an important challenge to nd ways to reduce the power dissipated in sensor data acquisition and transfer. One way to do this is by exploiting knowledge of the properties and uses of the transported data, to reduce communication power dissipation in exchange for, e.g., bounded inaccuracy of the communicated data (5–9). Reducing I/O power in exchange for bounded inaccuracy requires a physical channel property that provides such a tradeo . One ex0 0 1 0 1 0 1 0 0 0 1 0 0 0 0 0 0 0 1 0 1 0 1 0 0 0 1 0 1 0 0 1 Channel