The first big technical decision for an AI device is not the model. It is the compute tier, because it sets the cost, the battery life, the board size and what the product can do without a network.
The tiers
- Microcontroller class. Low cost and low power. Handles wake word, simple audio and sensors, and sends the heavy work to a phone or the cloud. Suits small, battery-powered products.
- Wireless SoC class. Adds radio and more memory, often running a real-time operating system. Suits audio devices, connected toys and wearables.
- Linux processor with an NPU. Runs vision, larger speech models and rich displays on the device, with a full operating system. Costs more in power, board area and bill of materials.
- Display and audio helper chips. Offload drawing, sound and motor control from the main chip.
What decides the choice
- What must work offline. A character that still responds without a network needs more local capability.
- Vision. A camera feeding a model usually pushes the product into the NPU tier.
- Display. A lively screen face needs more than a basic controller.
- Battery and size. Every step up costs power and space.
- Cloud cost per user. More on-device processing lowers it, which can change the business model.
Match tier to product
A desk companion with a screen and a camera sits in the upper tiers. A plush toy with speech sits lower. A wearable sits lower again. The right tier is the lowest one that does the job honestly, not the highest available.
Choose before the schematic
A later change of tier means a new board, new firmware and often new certification. Decide it from requirements, cost, supply and certification limits, not from the one chip a team happens to know.
We cover the four tiers, from NPU-class Linux processors to microcontrollers, on the capabilities page, and the cost trade-off in the AI toy core module article.
Planning an AI device? Tell us what you are building.