Power Consumption and Accuracy in Detecting Pedestrian Images on Neuromorphic Hardware Accelerated Embedded Systems

High-performance CPUs or GPUs have been used for the accuracy of AI applications to simulate a large number of neurons. However, the CPU and the GPU require high power consumption. Recently, possible adoptions of artificial intelligence in embedded systems have been considered because embedded systems are located at the edge to achieve faster response times and to reduce network load. However, CPUs and GPUs require too much power to achieve high performance in embedded systems that have limited power supply. To overcome this problem, employing special hardware for artificial intelligence has been studied. In this paper, we have implemented a pedestrian image detection system on an embedded device using NM500 neuromorphic chip. One NM500 chip contains 576 neurons, and we have measured the accuracy and power consumption, increasing the number of chips from one to seven. The results of the experiment show that the power consumption is linearly proportional to the number of neurons, while the accuracy is enhanced but not linearly proportional to the number of neurons. Based on the results of this experiment, we show that artificial intelligence hardware can be used to trade-off power consumption and accuracy in embedded systems.