Temperature compensation for analog machine learners

• Main Author: Tiong, Nicholas Kung Hung
• Other Authors: Arindam Basu
• Format: Final Year Project
• Language: English
• Published: 2019
• Online Access: http://hdl.handle.net/10356/77640
• Institution: Nanyang Technological University

Extensive use of Internet of Things (IoT) in daily lives has increased demand for ever more computational effort in cloud computing. However, with the aid of low-power machine learning system and edge analytics, part of the processing capability is delegated to edge devices closer to sensory front-end. Incorporation of such devices not only saved considerable amount of chip area, but also managed to cut down unnecessary power consumptions. The use analogue processing and extreme machine learning (ELM) algorithm was proved beneficial in designs of ultra-low power machine learners for “smart” sensors. However, the dependence of weights of these systems on the temperature cannot be easily ignored due to sub-threshold operation of transistors. This project aims to explore the best type of temperature behaviour for current reference in this system from behavioural simulations. Next, corresponding circuits will be designed and simulated using state-of-the-art IC design tools and CMOS processes. The paper also compares the temperature behaviour of current mode and voltage mode CMAs by using Physical Unclonable Function (PUF) metrics. Reliability of both modes under different temperature conditions is presented and alternative designs will be shown.