*1.2. Motivation and Contribution of This Study*

Smart edge devices rely on embedded systems with limited resources to process sensor data. For ML workloads requiring high computing power, energy-efficient systems are necessary. In recent years, many research efforts have been focused on energy-efficient embedded system designs to solve ML problems [5–9]. These systems are primarily designed to make inferences. However, embedded systems, especially at edge nodes, are likely to require online learning, control, and optimization capabilities. In autonomous cars, for example, using remote cloud-based training could lead to long communication delays. Therefore, embedded training devices are preferred. However, after being trained offline, ML inference models tend to diverge once in production. The retraining of such models will therefore require online training capabilities.

To understand how low-power embedded computing devices might help fill the aforementioned demand, this paper reviews the main design approaches for energy-efficient ML algorithm execution. In the following sections, it surveys candidates that meet both smart edge computing and Internet of Things requirements for low-power devices. CYSmart is a flexible and low-power smart edge computing system that we present as an example. A few working scenarios are used to evaluate the power efficiency of CYSmart.

### *1.3. Outline of the Paper*

First, we discuss energy-efficient computing systems dedicated to executing machine learning algorithms in Section 2. Section 3 provides a classification of low-power devices for IoT and smart edge computing on the basis of hardware resources and power dissipation constraints. This classification is then used to present a panorama of popular devices. The CYSmart low-power edge computing system is described in Section 4. Moreover, it is briefly compared with selected industrial edge computing technologies. Finally, some concluding remarks are provided in Section 5.
