Fine-Grained Power Gated Multiplier with Online Calibration for Medical IoT Devices

Abstract

With intensive research in the fields of machine learning and neural networks to improve its accuracy comes the responsibility to realize feasible hardware solutions on battery powered IoT devices. This work presents a study of analysis of power hungry computations and a fine-grained power gated multiplier design using approximation, that aims at energy optimization exploiting error resilience of these applications. We use truncation to reduce cycles and low power techniques to reduce power, thus achieving a 2-fold energy reduction. We use wearable IoT devices for medical purposes as our case study and show the generality of our work across applications. Our work performs similar to, or better than the latest work in the field and is a more generic implementation. We propose an online calibration mechanism to determine the approximation rate dynamically that maximizes energy optimization with very low accuracy loss. Our method uses a clustering solution to pre-determine the output label in a majority of cases, without having to need an inference model, thus further reducing energy. We achieve 78% energy improvement compared to a baseline implementation with just 0.46% accuracy loss across benchmarks.