Performance analysis of CNNs deployed on low power arquitectures

Abstract

Artificial intelligence is getting a more important role in our lives day by day. It is present in many daily used devices. In the majority of cases, the use of these algorithms implies having an internet connection so as to work due to their high computational demand. With the lately increase in the computing power of embedded systems, the execution of neural networks is focusing on the so-called edge computing. This favors the execution of neural network algorithms in environments where is not possible an internet connection. This article accomplishes an analysis of efficiency in the execution of a convolutional neural network (CNN) in a dedicated embedded device with graphical processing units (GPU) and inside two different multi-processors programmable system-on-chip with FPGA devices (MPSoC) from Xilinx. The first motivation for this work is the study of the viability of the deployment of an application to help visual functional diversity people with embedded systems not requiring an internet connection and oriented to Edge-AI, and the second is the search for a low-power device for longer battery life. At the end of the study, it is concluded that the deployment of these algorithms inside the selected low-power devices makes it possible to maintain a high frameper- second (FPS) rate in some architectures with an energy demand considerably inferior to the general-purpose devices.