A heterogeneous and reconfigurable embedded architecture for energy-efficient execution of convolutional neural networks

Abstract

Machine learning based convolutional neural networks (CNN) are becoming increasingly popular for identification tasks like image classification or speech recognition. However, CNNs have high memory and computational demands which makes it challenging to implement them on cost-efficient and energy-autonomous hardware. To cope with this challenge we present a heterogeneous and reconfigurable embedded architecture implemented on an inexpensive and widely available entry-level system on chip (SoC). Our architecture combines an ARM CPU and a coarse-grained reconfigurable architecture (CGRA) which execute a CNN in parallel to reach a higher energy-efficiency. Our results show up to 130% higher performance and 78% better energy-efficiency compared with an embedded Nvidia GPU.