Always-on Sensing

A sensor is a component that measures something from the environment: movement, sounds, brightness, or temperature. Normally, such a component is only active when you use the device. With always-on sensing, however, it runs continuously, even when the screen is dark. The English term literally means “always on”. So that this doesn’t drain the battery within a few hours, these sensors are coupled with a tiny, very low-power computing unit. This unit monitors the measured values and only wakes up the device’s main processor when something interesting happens.

Why devices prefer listening to waiting

Many useful features only work if a device doesn’t miss the moment. A smartwatch is supposed to detect a fall without you first opening an app. A voice assistant is supposed to react to its wake word even when the phone is in your pocket. Both are impossible without continuously running sensors, because the event is over before anyone could switch on the device.

The crucial point is power consumption. A smartphone’s main processor draws several watts under load. A specialized sensor chip often gets by with less than a thousandth of that. That’s why it can run for days on end while the big chip is almost always asleep. This division of labor is the actual trick behind always-on sensing.

At the same time, a conflict arises that regularly brings the technology into the news. A microphone that never turns off basically listens to everything. Manufacturers therefore emphasize that the raw data never leaves the device and that only the wake word is being searched for. Whether one believes this is a matter of trust that can hardly be verified technically from the outside.

The division of labor between small and large chip

The process consists of several stages. The sensor measures continuously, for example acceleration a hundred times per second. A very simple computing unit checks these numbers against a rough pattern. Only when the pattern matches does it switch on the next larger chip, which calculates more precisely. Experts call this principle a cascade, because the check becomes increasingly elaborate step by step.

At the lowest stage, tiny AI models often run today. Such a model is a program that has learned to recognize patterns from many examples. For always-on sensing, it must be shrunk down to a few kilobytes. This field is called TinyML, meaning machine learning on very small chips. Such models distinguish, for example, walking from climbing stairs, without any data going to the internet.

A common misconception is that the small stage must be especially reliable. The opposite is true: it’s perfectly fine for it to produce false alarms. What matters is only that it rarely misses genuine events. The second stage filters out the false alarms, and because it rarely kicks in, this costs hardly any energy.

From phone displays to industrial machinery

In everyday life, you constantly encounter always-on sensing without it being called that. The phone shows the time on the dark screen and rotates the image as soon as you tilt it. Fitness trackers count steps around the clock and measure your pulse in the background. Headphones notice that you’ve taken them out and stop the music. In all these cases, a sensor keeps running while the device appears to be off.

The technology is also widespread outside of consumer devices. In factories, sensors are attached to motors and listen for unusual vibrations to report defects early. In cars, they continuously monitor tire pressure and driving behavior. Such devices often run for years on a single coin cell battery.

In business news, the term usually appears in connection with chip manufacturers. Companies like Qualcomm, Bosch, or ARM advertise shifting more and more recognition directly onto the sensor. The technical term for this is edge computing: computing where the data is generated, instead of in a remote data center. For investors, this is interesting because such chips are built into devices in the billions.