Adaptive Approximate Computing on Hardware Accelerators Targeting Internet-of-Things

Abstract

Internet of Things (IoT) has been recognized as the new computing paradigm for connecting smart devices. These devices are typically power constrained and as their number and the volume of generated data increases, new techniques are required to reduce power consumption. Recent research has shown that approximate computing and Coarse-Grained Reconfigurable Arrays (GGRAs) are promising computing paradigms to reduce power consumption in a compute-intensive environment. CGRAs provide hardware acceleration and with the integration of approximate computing, there is substantial benefit in power and energy efficiency at the cost of arithmetic precision. In this paper, we present a methodology to support approximate computing on CGRAs, provide run-time error monitoring and control of distinct approximation levels on the CGRA. Experimental results show that CGRAs enhanced with approximate functional units provide more flexibility as they are able to adapt to error thresholds. Overall, with the proposed methodology we were able to satisfy specific error thresholds per evaluated use case and reduce power consumption up to 39% at best case.