Energy-efficient digital processing via Approximate Computing

Smart Systems applications often include error resilient computations, due to the presence of noisy input data, the lack of a unique golden output, etc. Therefore, computation accuracy constraints can be relaxed to improve a system's efficiency. Recently, a design paradigm called Approximate Computing (AC) has been proposed, that formalizes the exploitation of the accuracy dimension as a way to optimize efficiency in digital computing systems. AC configures configures as one of the most promising ways to reduce energy consumption in Smart Systems. In this chapter, we present an overview of the different AC techniques proposed in literature. Then, we focus on Algorithmic Noise Tolerance (ANT), one of the most suitable AC approaches for Smart Systems applications. In particular, we investigate for the first time the automatic application of this technique to an existing design. We show how this automation can be achieved with a flow that leverages standard EDA tools, with minimal input from the designer. Moreover, for a typical DSP circuit, we are able to obtain almost 45% total power saving.