Efficiency Limits for Value-Deviation-Bounded Approximate Communication

Abstract

Transferring data between integrated circuits accounts for a growing proportion of system power in wearable and mobile systems. The dynamic component of power dissipated in this data transfer can be reduced by reducing signal transitions. Techniques for reducing signal transitions on communication links have traditionally been targeted at parallel buses and can therefore not be applied when the transfer interfaces are serial buses. In this article, we address the issue of the best-case effectiveness of techniques to reduce signal transitions on serial buses, if these techniques also allow some error in the numeric interpretation of transmitted data. For many embedded applications, exchanging numeric accuracy for power reduction is a worthwhile tradeoff. We present a study of the efficiency of these value-deviation-bounded approximate serial data encoders (VDBS data encoders) and proofs of their properties. The bounds and proofs we present yield new insights into the best possible tradeoffs between dynamic power reduction and approximation error that can be achieved in practice. The insights are important regardless of whether actual practical VDBS data encoders are implemented in software or in hardware.