Power-compute co-design for robust pervasive IoT applications

Abstract

The modern development of internet of things (IoT) requires the IoT devices to be more compact and energy autonomous. Many of them require to be able to operate with unstable and low power supplies that come from various energy sources such as energy harvesters. This creates a challenge for building IoT devices that need to be robust to energy variations. In this research we propose methods for improving energy characteristics of IoT devices from the perspective of two main challenges: (i) improving the efficiency and stability of power regulators, and (ii) enhancing the energy robustness of the IoT devices. The existing design methods do not consider these two aspects holistically. One important feature of our approach is holistic use of event-based, temporal representation of data, which involves using asynchronous techniques and duty-cycle-based encoding. For power regulation we use switched-capacitor converters (SCC) because they offer compactness and ease of on-chip implementation. In this research we adapt the existing methods and develop new techniques for SCC design based on asynchronous circuits. This allows us to improve their performance and stability. We also investigate the methods of parasitic charge redistribution, and apply them to self-oscillating SCC, improving their performance. The key contribution within (i) is development of the methods of SCC design with improved characteristics. The majority of novel IoT systems are shifting towards the “AI at the edge” vision, for example, involving neural networks (NN). We consider a perceptron-based neural network as a typical IoT computing device. In our research we propose a novel NN design approach using the principle of pulse-width modulation (PWM). PWMencoded signals represent information with their duty cycle values which may be made independent of the voltages and frequencies of the carrier signals. As a result, the device is more robust to voltage variations, and, thus, the power regulation can be simplified. This is the second major contribution addressing challenge (ii). The advantages of the proposed methods are validated with simulations in the Cadence environment. The simulations demonstrate the operation of the designed power regulators, and the improvements of their efficiency. The simulations also demonstrate the principle of operation of the PWM-based perceptron and prove its power and frequency elasticity. The thesis gives future research directions into a deeper study of the holistic co-design of a variation-robust power-compute paradigm and its impact on developing future IoT applications.