Edge AI devices could make smart cities more energy efficient

Smart city systems are increasingly powered by AI operating across networks of Internet of Things (IoT) devices. These systems process vast amounts of data in real time to support applications such as activity recognition, infrastructure monitoring, and urban mobility analysis. While these technologies promise improved efficiency and decision-making, they also introduce new challenges related to power consumption and sustainable computing.

A new study, titled "Understanding Energy Efficiency of AI Deployments in IoT-Driven Smart Cities," published in IoT, examines how different hardware platforms perform when running AI workloads in edge computing environments and how these choices affect the overall energy efficiency of smart city deployments.

Hardware platforms shape the energy cost of smart city AI

Hardware architecture plays a decisive role in determining the energy efficiency of edge AI systems. Different platforms can vary dramatically in both power consumption and computational throughput, meaning that the same machine learning model can have vastly different energy footprints depending on where it is deployed.

At the lowest end of the power spectrum are intelligent sensor processors designed for ultra-low-power operation. These devices are capable of running simple machine learning models directly on sensor data while consuming extremely small amounts of energy. Such platforms are particularly suited for always-on monitoring tasks such as gesture recognition, motion detection, or environmental sensing. Because they operate within very tight energy budgets, they are ideal for battery-powered devices deployed throughout smart city environments.

Microcontroller-based systems represent another widely used platform for edge AI applications. Traditional microcontrollers offer moderate computing capability with relatively low power consumption. However, the study highlights that their performance improves significantly when combined with dedicated neural processing units. These specialized accelerators allow microcontroller platforms to run neural networks more efficiently while maintaining low energy use.

The research identifies these hybrid microcontroller systems as an important middle ground between ultra-low-power sensors and more powerful computing platforms. They provide improved inference performance while remaining suitable for energy-constrained environments. This balance makes them particularly attractive for edge intelligence tasks that require real time processing but cannot support the power demands of larger processors.

Field-programmable gate arrays represent another category examined in the study. These programmable hardware devices allow engineers to design customized computing architectures tailored to specific workloads. While they require more energy than ultra-low-power sensor processors, they offer advantages in terms of deterministic performance and system integration. For certain smart city applications that demand predictable response times or specialized hardware interfaces, FPGAs can provide an effective solution.

At the highest performance level are embedded graphics processing units. These platforms deliver significant computing power and are capable of running complex neural networks used in computer vision, object detection, and urban perception systems. Their high throughput allows them to support demanding applications such as autonomous driving assistance, real time surveillance analytics, and advanced environmental monitoring.

However, this performance comes at the cost of increased energy consumption. Embedded GPUs require substantially more power than other edge platforms, making them more suitable for installations where reliable power sources are available. For smart city deployments that require continuous high-performance AI processing, these systems provide the computational capacity necessary to handle large and complex models.

Matching AI workloads to the right hardware

Another key insight from the research is that no single hardware platform is universally optimal for all smart city AI tasks. Instead, the most efficient systems rely on carefully matching each application to the most appropriate computing platform.

For example, human activity recognition and gesture detection tasks often involve relatively lightweight machine learning models. These workloads can be handled efficiently by ultra-low-power sensor processors or microcontroller-based systems. Deploying such applications on more powerful hardware would unnecessarily increase energy consumption without delivering significant performance benefits.

Moderate edge intelligence workloads, such as sensor fusion or urban facility monitoring, typically require greater computational capacity. In these cases, microcontroller platforms equipped with neural processing units provide a strong balance between energy efficiency and performance. These systems can process data quickly enough for real time applications while still maintaining low power consumption.

More demanding computer vision workloads present a different set of requirements. Applications such as traffic monitoring, infrastructure inspection, and autonomous mobility support often rely on large neural networks capable of analyzing high-resolution images. Running these models efficiently requires the parallel processing capabilities of embedded GPUs or specialized hardware accelerators.

Selecting the wrong hardware platform can significantly increase energy consumption across large scale smart city deployments. For example, running simple machine learning models on high-performance GPU platforms may waste substantial amounts of energy, while attempting to execute complex neural networks on low-power devices may lead to unacceptable latency or performance limitations.

By carefully aligning hardware capabilities with application requirements, city planners and system designers can reduce energy consumption while maintaining the performance needed for real time urban services.

Toward sustainable AI Infrastructure in smart cities

Early smart city initiatives often focused primarily on functionality and connectivity. Today, sustainability and energy efficiency have become equally important considerations.

AI systems are increasingly deployed across large networks of devices that operate continuously. From traffic cameras and environmental sensors to connected public infrastructure, these systems generate vast streams of data that must be processed quickly and reliably. If energy efficiency is not considered during system design, the cumulative power consumption of these networks could become a major operational challenge.

The researchers argue that future smart city deployments should adopt a tiered hardware strategy that spans the entire sensor-to-edge computing continuum. Ultra-low-power devices can handle simple sensing tasks directly at the data source, reducing the need for continuous communication with central servers. Mid-range edge processors can perform intermediate analytics and decision making, while high-performance platforms handle computationally intensive workloads.

This layered approach allows cities to distribute computing tasks efficiently across multiple device classes. Instead of relying on a single hardware architecture, smart city systems can allocate workloads to the most appropriate platform based on energy constraints and computational requirements.

The study also highlights the importance of software optimization and model design in improving energy efficiency. Techniques such as model quantization, pruning, and lightweight neural network architectures can significantly reduce computational overhead without sacrificing accuracy. Combining these software strategies with energy-efficient hardware platforms can further enhance the sustainability of AI deployments.

Another important contribution of the research is the development of a new dataset designed for evaluating embedded AI systems. This dataset focuses on gesture recognition tasks and provides a benchmark for testing machine learning models on low-power sensor platforms. By supporting reproducible experiments, the dataset allows researchers to better compare the performance of different hardware architectures in energy-constrained environments.

The researchers suggest that future work should expand this approach to include additional datasets and real-world deployment scenarios. Long-term experiments involving operational smart city systems could provide deeper insights into how hardware platforms perform under continuous workloads and varying environmental conditions.