*4.4. Gained Insights and Discussion*

As a representative ultra-low-power device of the CYSmart system, the CYComs were the primary focus of the above use cases. There is also another component, the CYEdge, which embeds a Raspberry Pi 4 and a LoRa communication shield. Based on the power measurements of the CYEdge under normal operating conditions, it can be classified as a class 3 device, as described in Table 1. Below are some insights regarding CYSmart's current implementation and potential improvements.

**Gained insights.** The CYCom component of the CYSmart system utilizes a commercialoff-the-shelf (COTS) microcontroller manufactured by STMicroelectronics. Choosing this approach reduces the cost of the component as well as the development time. A CYCom's CPU is the primary energy consumer in the aforementioned use cases, followed by the LoRa module and the sensor power supply. Each of these three modules can be improved.

STM32 boards are based on von Neumann microarchitecture, leading to costly data movement between different hardware units. As a result, future improvements could include designing a customized solution that meets the requirements of the domain applications. This is consistent with the notion of domain-specific hardware accelerators as described in [85]. There is a lot of power consumed during one routine iteration in the second use case from the previous section without any data processing being performed. This unnecessary power consumption must be eliminated in order to improve the energy efficiency of the system. This problem may be solved by means of power gating, for example. A customized solution that incorporates such a mechanism is therefore desirable. Suitable design approaches should be considered for design space exploration by selecting high-level methodologies, e.g., [86–88], covering different abstraction levels: high-level analytical modeling [89–92], transaction-level modeling [93–95], cycle-accurate design [96–98], or register transfer level [99]. As surveyed in this paper, it is possible to implement the architecture using FPGA or ASIC designs at the expense of higher costly implementation efforts. As for CYComs, the CYEdge power consumption can be reduced by applying the same design methods.

Sensor power consumption is difficult to reduce since it is heavily dependent on the type of sensor being used. A wide range of digital and analog sensors can be interfaced with the CYCom. External 24 V lithium-ion batteries are currently used to power the integrated sensors. Instead of analog sensors, digital sensors with internal 3.3 V batteries could be considered here to reduce power consumption. Depending on the measurement environment, LoRa modules consume varying amounts of power. In both Cases I and II, the system can communicate across a reinforced concrete wall 90 cm thick with the initial parameters. In the case of a 15 dBm data transmission capacity and a spreading factor of 12, the maximum transmission delay is 2 s. Therefore, its maximum power consumption is 166 mW. These parameters can be adjusted according to the operational environment in order to reduce the LoRa module's power consumption.

**Comparison of CYSmart w.r.t. selected industrial solutions.** As a mature low-power edge computing solution, the CYSmart system can be compared with a number of industrial technologies. For this purpose, we consider some relevant criteria, described as follows:


In light of the aforementioned criteria, Table 6 provides a comparison of CYSmart w.r.t. the industrial edge computing technologies summarized in the sequel. The *TMI-Orion* company proposes a solution for the design and manufacture of high level technologies that target harsh environments. A key component of its edge computing technology is a network of smart sensors such as NanoVACQ Fullradio [100], which communicate via a 2.4 GHz radio protocol with a Radio Transceiver [101]. Using a serial protocol, the latter transmits data to a host computer that manages and displays data from a sensor network. The *Gravio* company develops an IoT platform that is capable of connecting several sensors. Using the ZigBee wireless protocol, these sensors communicate with the Gravio Hub [102]. Data can be viewed and managed by users.

The *moneo appliance* is an edge solution manufactured by IFM company [103]. It consists of a dedicated software toolbox that allows for the management of sensor parameters as well as data display. Templates are provided in the toolbox for defining network configurations. Sensors are connected to the moneo appliance via an IO-Link Master, which serves as an interface between the appliance and the gateway computer.

The *Advantech* company developed an IoT solution that relies on data measurement devices named WISE (e.g., WISE-4060 [104]) and an intelligent edge server (e.g., EIS-D150 [105]). By using the WiFi protocol, the WISE devices send data from the sensors to the edge server. Users are provided with a real-time dashboard for managing WISE devices. The *InHand Networks* company has defined a specific gateway [106], which provides data optimization in the IoT infrastructure and provides real-time response times. The gateway device can be connected to a local network. It is compatible with real-time Ethernet protocols and supports the Docker software system.

The MCM200 series components (e.g., MCM-204 [107]) are edge computing solutions designed by the *Adlink* company. They are standalone data acquisition devices (i.e., no host computer is required) that can monitor, analyze, and execute real-time actions. WiFi or Ethernet ports are available for communication.

Finally, the *Analog Devices* company offers the *SmartMesh Wireless HART* technology which consists of a small network manager (LTP5903-WHR [108]) that communicates with a number of sensor nodes called "motes" (e.g., LTP5900-WHR [109]). The network manager and motes must be programmed by the user. The network manager is responsible for centralizing data and communicating it to the host computer. Using analog data from sensors, the motes transmit data to the network manager.

Table 6 globally illustrates that CYSmart and Advantech technologies offer several advantages over other solutions. There are many similarities between these two technologies; however, CYSmart is capable of supporting a larger wireless communication distance than Advantech's solution. Because of this, CYSmart is well suited to critical environments, such as nuclear power plants.


*J. Low Power Electron. Appl.* **2022**, *12*, 61

**Table 6.**

Comparison

 of CYSmart with similar edge computing

technologies.

#### **5. Summary**

Embedded architectures for future edge devices likely will need to support training, control, and optimization capabilities, according to the current trends in edge computing. In this paper, we discuss recent efforts regarding energy-efficient hardware solutions for machine learning at the edge. We reviewed current design approaches and devices targeted at implementing IoT and smart edge computing with limited computing and power capabilities. Candidate low-power devices that could meet IoT and smart edge computing requirements have been surveyed. CYSmart, a flexible low-power edge computing system, was demonstrated as an interesting solution. To evaluate its power efficiency, a few working scenarios have been considered. Finally, a brief comparison of CYSmart with selected industrial edge computing technologies was presented.

**Author Contributions:** Conceptualization, L.M.W., J.-M.B., G.B. and A.G.; methodology, L.M.W., J.-M.B. and A.G.; software, L.M.W. and J.-M.B.; validation, L.M.W., J.-M.B. and A.G.; writing—original draft preparation, L.M.W., J.-M.B., G.B. and A.G.; writing—review and editing, L.M.W. and A.G.; supervision, J.-M.B., G.B. and A.G. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Acknowledgments:** The authors would like to thank Guillaume Devic and Gilles Sassatelli for their feedback in early discussions on part of the current work.

**Conflicts of Interest:** The authors declare no conflict of interest.

## **Abbreviations**

The following abbreviations are used in this manuscript:

