Energy and error analysis framework for approximate computing in mobile applications

In this brief, we propose a framework that enables us to analyze energy and error for mobile applications when run on systems with approximate circuits. Approximate circuits have conventionally been used in image/video processing applications, which are mostly limited to 8-bits. To the best of our knowledge this is the first work where approximate circuits have been evaluated on a 32-bit processor running real mobile applications. We observe that in approximate adders where Carry is approximated, with 2-bit approximation in LSB, the absolute average error in image processing applications is 3. However, in mobile applications approximation in Carry can lead to an overflow. Hence, approximate adders with Carry approximation are not suitable for mobile computing. We also show the role of data-dependent switching in energy consumption and highlight which input pattern should not be approximated to obtain lesser error. In this brief, we also propose design of three energy-efficient approximate hybrid CMOS full-adders with varying levels of inaccuracies. The adder designs are implemented in UMC 65-nm technology using Cadence Virtuoso. Compared to existing approximate adders, on an average, the proposed adders consume 44% lesser energy and have 2 × lesser energy delay product. Our proposed adder designs have similar leakage power as compared to the existing adders.