The Design of Efficient Data Flow and Low-Complexity Architecture for a Highly Configurable CNN Accelerator

This paper presents a highly configurable and low-complexity CNN accelerator based on the MobileNetV3 model. To the best of authors’ knowledge, this is the first design of CNN accelerator based on the MobileNetV3 model. A highly efficient processing flow and memory-access scheme are proposed in this paper so that the throughput is greatly enhanced for the structural features in MobileNetV3 model. Furthermore, the proposed processing flow enhances the efficiency for the utilization of hardware components to reduce the complexity. Based on the proposed processing flow, this paper presents a highly configurable architecture to support various operation modes in MobileNetV3 model. The designed architecture is synthesized and layout with TSMC 90 nm technology. The evaluations for the performance and area complexity are conducted based on the post-layout estimations. It is shown in this paper that the performance of 197.7 FPS is achieved with the hardware complexity of 5392 KGEs for the MobileNetV3-Large. Compared to the state-of-the-art accelerator based on MobileNet, the FPS of the proposed design is improved by 3.4 × and the complexity is reduced by 18%.