Topic Questions
Where do you see Edge AI doing a good job and falling short?
-
Edge AI becomes particularly valuable when traditional rule-based algorithms struggle with the messiness of real-world environments. In audio detection, for example, the same event can look different depending on distance, ambient noise, and other conditions, making fixed thresholds prone to false positives. Machine learning can handle this variability by learning patterns across multiple features and over time, improving accuracy and efficiency.
-
The bigger challenge is making AI practical on embedded devices. Models need representative training data, which can be difficult to collect and label across different environments. There is also more to manage than simply running the model, including training, validation, updates, and long-term maintenance. And unlike traditional algorithms, machine learning models can be harder to predict and debug - an important consideration when reliability is critical.
-
That doesn’t mean Edge AI needs to replace traditional algorithms. In many cases, the two work together: deterministic processing such as filtering and feature extraction prepares the data, while AI handles more complex patterns. This approach can keep models smaller while tackling problems that traditional methods struggle with, such as image recognition and sensor fusion. The goal isn’t to replace everything with AI, but to use it where it adds the most value while keeping the rest of the system simple and predictable.
What are some common stumbling points as AI gets deployed at the edge?
-
The move toward neural networks at the edge has largely come from the limitations of rule-based systems, but it has introduced new challenges of its own. As AI workloads become more memory-intensive, edge devices can be constrained by available memory, power, and processing hardware. Models often need to be compressed and quantized to fit within those limits, sometimes affecting performance compared with what was achieved during development.
-
AI also has to work within the larger embedded system, where power management, wireless communication, timing, and scheduling can all influence performance. Keeping models updated as conditions change adds another layer of complexity.
-
But hardware may not be the biggest obstacle. Getting the right data can be even more difficult. While applications like smartphones can justify significant investment because the same model reaches millions of devices, edge applications such as motor controls and appliances can involve thousands of unique systems, each operating differently. Building enough high-quality, representative data for those environments remains one of the biggest challenges - and one that can’t simply be solved with more computing power.
What Edge AI applications are currently production ready, where do you see future applications going?
-
At the extreme edge, the focus is less on what AI can do and more on how much processing can be moved directly onto the sensor while using as little energy as possible. Wearables are a good example. Step counting, for instance, has evolved from simple thresholds that were easy to fool into machine learning models that can much more accurately distinguish real activity from false signals.
-
The same approach is being applied to vibration sensing, predictive maintenance, asset tracking, and industrial fault detection. By making decisions locally, such as identifying an anomaly before sending anything to the cloud, devices can reduce bandwidth and energy consumption. Other applications include vision-based quality inspection and cameras that can distinguish between a family member, a pet, or an intruder.
-
As edge hardware becomes more capable and power-efficient, these applications are moving from proof of concept into everyday products. The goal isn’t necessarily to make AI more powerful, but to push useful intelligence as close to the source of the data as possible.
-
Voice assistants illustrate why that matters. Instead of continuously sending audio to the cloud, local AI can handle tasks such as wake-word detection directly on the device. Moving more processing to the edge can reduce power and bandwidth requirements while also keeping sensitive data, such as conversations inside the home, from leaving the device. In that sense, Edge AI can make intelligent devices more practical, private, and efficient simply by allowing them to make more decisions locally.
Can embedded engineers work with nondeterministic systems, like Edge AI or is there a clear division?
-
Development toolkits are an important part of making Edge AI practical, but they are only one piece of the puzzle. Teams still need to clearly define the problem, understand their data, and determine what level of accuracy the application actually requires. The right tools can then bring together everything from model compression and compilers to specialized processors and ML frameworks to turn an idea into a working system.
-
That process also requires embedded engineers to expand beyond the deterministic mindset they’re used to. Traditional development focuses on precise timing and predictable behavior, while AI introduces probability and uncertainty. Engineers need to become comfortable asking not only “Does it work?” but “How accurate does it need to be?”
-
The answer depends on the application. A 95% accurate people counter might be perfectly acceptable, while that same accuracy would be unacceptable for a safety-critical device. Successful Edge AI isn’t about replacing deterministic engineering; it’s about combining it with machine learning and choosing the right level of precision for the problem.
Do Software Development Kits accelerate development or encourage vendor lock-in?
-
SDKs are helping make Edge AI more accessible to traditional embedded engineers by providing features such as AutoML and libraries of pre-trained models. Rather than requiring every engineer to become a data scientist, these tools can shorten the path from an idea to a deployed AI application. As AI also becomes more common in engineering education, the embedded engineer’s skill set is evolving alongside the technology.
-
SDKs can also improve portability when they build on open-source frameworks and standards. However, vendor lock-in hasn’t disappeared. Many platforms still rely on proprietary SDKs, compilers, and deployment pipelines underneath those abstraction layers. Open frameworks can make applications easier to move between hardware platforms, but they can also introduce additional resource requirements and complexity.
-
Ultimately, engineers still face a trade-off: use a vendor’s tools for maximum performance and efficiency, or use higher-level, more portable frameworks for convenience and flexibility. The goal is greater interoperability, where a model could move between platforms with minimal changes, but in the resource-constrained embedded world, portability often comes with its own costs.
YouTube Video Sources
If you would like to watch this video or any other Let’s Talk Technical Series, please utilize the sources below.
Video on this Topic
Let’s Talk Technical: Edge AI Part 1, Theory to Deployment with Industry Experts | DigiKey
Other Videos from this Series
Let’s Talk Technical Playlist - YouTube
Suppliers Featured in Video
NXP Semiconductors
NXP enables secure connections and infrastructure for a smarter world, advancing solutions that make lives easier, better and safer. As the world leader in secure connectivity solutions for embedded applications, NXP is driving innovation in the secure connected vehicle, end-to-end security and privacy and smart connected solutions markets.
STMicroelectronics
STMicroelectronics is a global independent semiconductor company and is a leader in developing and delivering semiconductor solutions across the spectrum of microelectronics applications. An unrivaled combination of silicon and system expertise, manufacturing strength, Intellectual Property (IP) portfolio and strategic partners positions the Company at the forefront of System-on-Chip (SoC) technology and its products play a key role in enabling today’s convergence trends.
Microchip Technology
Microchip Technology Inc. is a leading provider of microcontroller and analog semiconductors, providing low-risk product development, lower total system cost and faster time to market for thousands of diverse customer applications worldwide. Headquartered in Chandler, Arizona, Microchip offers outstanding technical support along with dependable delivery and quality.
Silicon Laboratories
Silicon Labs is the leading innovator in low-power wireless connectivity, building embedded technology that connects devices and improves lives. Merging cutting-edge technology into the world’s most highly integrated SoCs, Silicon Labs provides device makers with the solutions, support, and ecosystems needed to create advanced edge connectivity applications. Headquartered in Austin, Texas, Silicon Labs has operations in over 16 countries and is the trusted partner for innovative solutions in the smart home, industrial IoT, and smart cities markets.
Nordic Semiconductor
Nordic Semiconductor is a fabless semiconductor company specializing in short-range wireless and low-power cellular IoT. Nordic pioneered ultra-low-power wireless and helped develop Bluetooth Low Energy, a widely-adopted wireless technology. The company’s award-winning Bluetooth LE solutions have made it the market leader, and are complemented by ANT+, Thread, and Zigbee products. Nordic’s reputation is built on the supply of leading-edge wireless technologies and development tools that shield designers from RF complexity. This approach extends to Nordic’s latest technology, cellular IoT. Launched in 2018 after four years of development, this NB-IoT and LTE-M solution leverages cellular infrastructure to extend the IoT.