Reducing serial I/O power in error-tolerant applications by efficient lossy encoding

Abstract

Transferring data between integrated circuits (ICs) accounts for an important fraction of the power dissipation in wearable and mobile systems. Reducing signal transitions reduces the dynamic power dissipated in the data transfer between ICs. Techniques such as Gray coding to reduce transitions between two parallel words cannot be applied when the signal transitions are between bits of a single serialized word. This paper introduces value-deviation-bounded serial encoding (VDBS encoding). VDBS encoding significantly reduces signal transitions between bits of a single serialized word, trading power efficiency for data accuracy. This tradeoff is worthwhile when the data are from signal sources such as sensors and destined for consumption by signal processing algorithms that are error-tolerant. We present analytic formulas for the Pareto-optimal VDBS encoders and introduce an efficient algorithm, Rake, that reduces transitions almost as much as the optimum transition-reducing encoder. We evaluate Rake by encoding data in a pedometer system and in a text-recognition system. For the pedometer, Rake reduces transitions by 54% on average, in exchange for step count errors smaller than 5%. For the text recognizer, Rake reduces transitions by 55% on average, while maintaining OCR accuracy above 90 %.