INVITED PAPER: An Energy-Efficient and Reconfigurable DNN Accelerator for Optic-Fiber based Edge Sensing and Computing.

The signals of distributed optic-fiber may represent information related to certain events which can be classified by Deep Neural Networks (DNN). However, energy-efficient and real-time classification with different DNNs for different edge applications is a great challenge. This paper proposes a reconfigurable hardware accelerator design for DNNs with different structures to classify the signals from various optic-fiber sensors efficiently in real time. By exploiting the arithmetic operation similarity, the proposed reconfigurable design reuses a PE array to implement different convolutional layers and network structures of DNNs, and also utilizes sparsity in DNNs to achieve high energy efficiency and low computation latency. In addition, a design space exploration method is proposed to reach an optimal balance between mitigating on-chip memory overhead and minimizing computation latency so as to effectively achieve low latency and power consumption. FPGA implementation shows that it can achieve a reconfigurable design with 26484 LUTs and 14366 FFs. It outperforms the CPU/GPU solution, i.e., 137 and 99 times in energy efficiency at 180 MHz, respectively. A preliminary Python simulation results in a case study shows that the accuracy of Resnet-12 implemented by the proposed design is 98.94% on the optic-fiber vibration signal classification task.