A Power-Efficient Programmable DCNN Processor for Intelligent Sensing

Abstract

Existing deep convolutional neural network (DCNN) processors are mainly designed for high-end applications such as autonomous vehicle, data center and smart phone where the design focus is the performance, while for intelligent sensing devices power efficiency are more important. In addition, programmability is important for DCNN processors to support different DCNN. We have proposed a power-efficient programmable DCNN processor dedicated for intelligent sensing devices and demonstrated it using FPGA. Several techniques have been proposed to improve the power efficiency. Implemented on a Xilinx VC707 FPGA board, It achieves a power efficiency of 31 Gops/W with peak performance of 487 Gops, which is better than several state-of-the-art DCNN processors.