Workflow on CNN utilization and inference in FPGA for embedded applications: 6th South-East Europe Design Automation, Computer Engineering, Computer Networks and Social Media Conference (SEEDA-CECNSM 2021)

The operation of devices such as wearables, hearables, mobile telephones and their accessories can be challenging due to the limitations of embedded systems restrictions in size and computational power. Design scalability and flexibility are necessary factors to better support the variation in utilization of applications. Convolutional Neural Networks (CNN) are able to address the plurality of requirements with their specialization in multi-dimensional, grid like topologies. At the time of proceeding with this study, Field Programmable Gate Arrays (FPGA) have demonstrated as most suitable for latency-sensitive, hardware acceleration and real-time inference jobs. In this work, we present the workflow utilizing a custom CNN model that is small enough to fit and operate in low cost FPGA, such as the Zynq7020 device available to the ZedBoard. The end to end steps to reproduce this work are available at https://github.com/ECSAlab/dnndk-zedboard.