An Energy-Efficient and Flexible Accelerator based on Reconfigurable Computing for Multiple Deep Convolutional Neural Networks

Multiple Convolutional Neural Networks (CNNs) become widely used in modern AI systems. There is increasingly necessity to apply different CNN shapes for different scenarios. However, it also brings challenges on throughput, energy efficiency and flexibility to hardware. In this paper, a novel accelerator, called reconfigurable neural accelerator (RNA), was proposed based on reconfigurable computing technology. In addition, image row broadcast (IRB) and zero detection technology (ZDT) were applied for increased energy efficiency and throughput. IRB can optimize the convolutional dataflow on spatial array architecture with 22×22 processing elements, increasing data reuse and reducing data movement. ZDT reduces the weight data access of the fully connected layer. At the cost of 10.25W power consumption on Virtex UltraScale XCVU440 platform, RNA can process the convolutional layers at 97.4 GOPS for AlexNet, at 90.75GOPS for VGG and at 100.8 GOPS for Lenet-5, respectively.