When Does Edge AI Make Sense on an Embedded Device, and What Hardware Does It Need?

Running the model on the device is not a fashion; it is a constraint decision. Here is how to make it, and what it means for the board.

· by Kari Rantakoski

Short answer: Edge AI, running inference on the device itself, makes sense when the decision must be made in milliseconds, when connectivity is absent, expensive or unreliable, when raw data must stay on site, or when cloud data volume would cost more than the device. Cloud AI remains right for training, fleet-wide analytics and models too large for embedded hardware. Most industrial systems end up hybrid: inference at the edge, training and monitoring in the cloud. On the hardware side, edge AI spans three tiers: microcontroller inference for simple classification and anomaly detection, application processors with NPUs for vision and multi-sensor fusion, and dedicated accelerators for camera-rate deep learning.

Decision criteria

CriterionFavours edgeFavours cloud
LatencyControl loops, safety reactions, under 10–100 msSeconds are acceptable
ConnectivityRemote sites, mobile machines, metered linksReliable wired network
DataPrivacy, high-bandwidth sensors (vibration, camera)Low-rate telemetry
CostMany devices, long life, data transfer priced per MBFew devices, short life
ModelStable, small, quantisedLarge, frequently retrained
OperationsMust work offlineCentral updates preferred

Hardware tiers for edge inference

  • Tier 1, microcontroller: ARM Cortex-M4/M7/M33 with 256 KB–2 MB flash running quantised models via TensorFlow Lite Micro or vendor toolchains. Vibration, current signature and acoustic anomaly detection, simple classification. Milliwatts, no heat sink.
  • Tier 2, application processor with NPU: NXP i.MX 8M Plus class, Cortex-A with 1–3 TOPS neural units, embedded Linux. Camera-based inspection, multi-sensor fusion, OPC UA and gateway roles. A few watts.
  • Tier 3, accelerator: dedicated NPUs or GPU modules for multiple camera streams and large models. Tens of watts, thermal design and power supply become the main board design task.

What edge AI does to the PCB

The model is software, but the board decides whether it runs: sufficient RAM for activations, fast external flash for model storage, clean sensor front-ends with correct sampling rates and anti-aliasing, power supply headroom for inference bursts, thermal paths for NPUs and a secure element for model and data protection. These are schematic and layout decisions taken in stage 2 of the product development process, and they are why Comtec Labs designs the AI hardware and the sensor electronics together.

Key facts

  • Quantised 8-bit models routinely run anomaly detection on a Cortex-M7 in a few milliseconds.
  • An NPU-equipped application processor handles camera inspection at a few watts, where a GPU module needs tens.
  • Data transfer, not compute, is often the largest running cost of a cloud-only design with high-rate sensors.
  • Comtec Labs designs, prototypes and manufactures edge AI electronics, sensor front-ends and the firmware that runs the model.

How we work on an edge AI project

Feasibility on an evaluation board with your data first, then a sensor and compute architecture, a custom PCB with the right tier of processor, firmware with the inference runtime, model deployment and update pipeline, and production with programming and calibration on the line. Applications include production-line monitoring, energy infrastructure, predictive maintenance and safety systems.

Frequently asked questions

Can a microcontroller really run AI?

Yes for small, quantised models: anomaly detection, keyword spotting, simple classification. Vision and larger networks need an application processor with an NPU or an accelerator.

How do we update the model in the field?

Through the same signed over-the-air update mechanism as the firmware, with the model stored separately so it can be updated without a full firmware release.

Does Comtec Labs train the models?

We integrate and deploy models; training is usually done by the customer or a data science partner with data from the prototypes we build. We help design the data collection.

What about power and heat?

Tier 1 designs run on milliwatts; tier 2 needs a few watts and a thermal path; tier 3 needs active cooling and a proper power supply. We design the board for the chosen tier.

Ready to reduce PCB surprises?

Comtec Labs offers a full suite of services to streamline your workflow:

PCB design service
PCB prototyping service
PCB component sourcing
PCB component assembly
PCB testing service
PCB repair and modifications
Printed circuit board production
PCB mass production

← Comtec Labs blog

Ready for your next project?

From R&D to scalable manufacturing.

Contact us