Design of a Generic Dynamically Reconfigurable Convolutional Neural Network Accelerator with Optimal Balance

Abstract

In many scenarios, edge devices perform computations for applications such as target detection and tracking, multimodal sensor fusion, low-light image enhancement, and image segmentation. There is an increasing trend of deploying and running multiple different network models on one hardware platform, but there is a lack of generic acceleration architectures that support standard convolution (CONV), depthwise separable CONV, and deconvolution (DeCONV) layers in such complex scenarios. In response, this paper proposes a more versatile dynamically reconfigurable CNN accelerator with a highly unified computing scheme. The proposed design, which is compatible with standard CNNs, lightweight CNNs, and CNNs with DeCONV layers, further improves the resource utilization and reduces the gap of efficiency when deploying different models. Thus, the hardware balance during the alternating execution of multiple models is enhanced. Compared to a state-of-the-art CNN accelerator, Xilinx DPU B4096, our optimized architecture achieves resource utilization improvements of 1.08× for VGG16 and 1.77× for MobileNetV1 in inference tasks on the Xilinx ZCU102 platform. The resource utilization and efficiency degradation between these two models are reduced to 59.6% and 63.7%, respectively. Furthermore, the proposed architecture can properly run DeCONV layers and demonstrates good performance.