Reducing Power Consumption using Approximate Encoding for CNN Accelerators at the Edge

Abstract

Convolutional neural networks (CNNs) have demonstrated significant potential across a range of applications due to their superior accuracy. Edge inference, in which inference is performed locally in embedded systems with limited power resources, is researched for its energy efficiency. An approximate encoder is proposed in this study for decreasing switching activity, which minimizes power consumption in CNN accelerators at the edge. The proposed encoder performs approximate encoding based on a pattern matching of a comparison pattern and current data. Software determines the value of the comparison pattern and the availability of the recommended encoder. Experiments with a CIFAR-10 dataset utilizing LeNet5 show that using the suggested encoder, depending upon the comparison pattern, power consumption of a CNN accelerator can be reduced by 21.5% with 1.59% degradation on inference quality.