A Low Power Artificial Intelligence Chip for Electric System Edge Applications

Abstract

In recent years, artificial intelligence especially deep learning technology has made rapid progress. Its ability in image recognition and target detection has been close to or even exceeded than that of human beings.  In the electric power system, decision making and assistant decision making based on artificial intelligence technology is applied in more and more scenarios. Among them, video surveillance is one of the most practical scenarios. In this paper, aiming at the intelligent demand of edge measuring equipment, take the online monitoring of the state of the transmission line as the target application scenario, designed a low-power artificial intelligence reasoning chip based on "near memory operation", with the computing unit as the core, the high power problem caused by the "memory wall" can be reduced by distributed nearby deployment of weights, parameters and hardware operators, so as to greatly reduce the power consumption of the chip and improve the chip operation efficiency. The experimental results show that the artificial intelligence chip based on this architecture has a power consumption of less than 0.5W and a recognition throughput of more than 50FPS, which can effectively improve the local perception ability of the data of the side devices and facilitate the safe and stable operation of the power grid.