Date of Graduation
Bachelor of Science in Computer Engineering
Computer Science and Computer Engineering
Committee Member/Second Reader
The Internet of Things (IoT) consists of all devices connected to the internet, including battery-powered devices like surveillance cameras and smart watches. IoT devices are often idle, making leakage power a crucial design constraint. Currently, there are only a few low-power application-specific processors for deep learning. Recently, the Trustable Logic Circuit Design (TruLogic) laboratory at the UofA designed an asynchronous Convolutional Neural Network (CNN) system. However, the original design suffered from delay-sensitivity issues undermining its reliable operation. The aim of this thesis research is to modify the existing CNN circuit to achieve increased reliability and to optimize the improved design for low-power, IoT applications. Simulations demonstrate that the delay-sensitivity modifications allow the CNN system to operate more reliably. The optimized CNN circuit consumes 9.9% less active energy and 14% less leakage power. In addition, the optimized CNN circuit requires 8.24% smaller area and is 1.43% faster than the more reliable CNN circuit. These improvements show the proposed CNN system is better suited for IoT applications.
Sherrill, C. (2019). Optimization of Ultra-Low Power Application-Specific Asynchronous Deep Learning Integrated Circuit Design. Computer Science and Computer Engineering Undergraduate Honors Theses Retrieved from https://scholarworks.uark.edu/csceuht/67