**1. Introduction**

The edge computing paradigm [1] is an emerging paradigm for Internet-of-Things systems where computations are distributed across a variety of compact devices in order to bring computing capability closer to data sources, such as environmental sensors and cameras. We can mention the following advantages of edge computing over the traditional centralized computing paradigm found in cloud systems:


Figure 1 illustrates the hierarchical layers of a typical edge computing infrastructure. Sensors in Layer 0 collect data from the environment first. In subsequent layers, the data are processed with appropriate devices based on the complexity of the processing. To meet the edge computing requirements, devices are placed close to sensors.

**Citation:** Martin Wisniewski, L.; Bec, J.-M.; Boguszewski, G.; Gamatié, A. Hardware Solutions for Low-Power Smart Edge Computing. *J. Low Power Electron. Appl.* **2022**, *12*, 61. https://doi.org/10.3390/ jlpea12040061

Academic Editor: Orazio Aiello

Received: 28 October 2022 Accepted: 21 November 2022 Published: 25 November 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

**Figure 1.** Hierarchical smart edge computing.

#### *1.1. Machine Learning at the Edge*

For edge computing nodes, implementing data analytics, particularly machine learning (ML), is a major challenge. Applications that leverage ML techniques at the edge are numerous. In essence, they deal with the inference problem. Real-time video analytics is a prominent application found in systems such as video surveillance, traffic control, and autonomous vehicles. Another notable application is feature extraction from images, for example, detecting areas and objects, identifying handwritten characters, and monitoring healthcare. To ensure people's safety (for instance, in the event of a fire) or to minimize energy consumption by utilizing renewable resources, smart homes and cities incorporate devices that use ML techniques for sensing and controlling the environment. Amazon's Alexa has made automatic speech recognition popular due to its success.

Based on the richness of ML techniques, three main families can be distinguished [2]. In *supervised* learning, classification and regression tasks are often achieved with algorithms such as support vector machines (SVM), artificial neural networks (ANNs), and linear regression. With the use of algorithms like k-means or x-means, *unsupervised* learning can be used for clustering and prediction tasks. *Reinforcement* learning focuses on decision-making through Q-learning. The majority of machine learning algorithms implemented on edge devices currently utilize *inference* (i.e., the process of directly solving ML problems with pre-trained ANNs) rather than *training* (i.e., the process of minimizing error as a function of ANN parameters, given an ML problem). One reason cited in [3] is the high bandwidth and latency costs involved in exchanging network updates across multiple edge devices (centralized model training might be more efficient since updated networks would be transferred directly to devices). Another concern in the design of edge nodes is energy and hardware costs. Structured labeled data is usually used for inference. The case is different for *deep learning* [4], which is used for training tasks in which precision is critical. In this process, complex multi-layer ANNs are applied along with huge amounts of raw data. Edge devices lack the computing power and data storage capacity needed for this.
