TinyML chips are microcontrollers and other tiny, low-power chips with a small AI engine, which run a machine learning model inside the device itself, with no cloud. TinyML is short for tiny machine learning. The models are small: often tens to a few hundred kilobytes. The chips use a few thousandths of a watt or less while they work, so a device can run for years on a battery. Some cost less than ₹100.
Why run the model on the device? Sending raw sensor data to the cloud uses battery power and network data, and it takes time. Some data should not leave the device at all, such as a camera's pictures. If the model runs on the chip, the device sends only the answer: 'normal', 'a person entered', or 'a bearing is wearing out'. Running computing where the data is collected is called edge computing.
The chip only runs the model, a job called inference; it does not train it. The difference is explained on training vs inference. Engineers train the model on servers, using recordings from many machines. Then they shrink it, for example by storing each number in 8 bits instead of 32, and load it into the chip's flash memory. The tools for building and shrinking such models are described under edge AI. A model is mostly many multiply-and-add sums. The chip's AI engine does many of these sums at once, using much less energy than the microcontroller's ordinary processor core would.
Consider a textile mill in Tiruppur with 120 motors driving its spinning machines. A sensor about the size of a matchbox is fixed to each motor, with a battery meant to last for years. Every 10 minutes, the sensor wakes up and records 1 second of vibration from its accelerometer, at 4,000 readings a second. Its TinyML chip then runs a model of about 100 KB, trained on vibration recordings of healthy and worn bearings. The model takes a few thousandths of a second. If the answer is 'normal', the sensor sends nothing and goes back to sleep. One week, the bearing in motor 37 starts to wear, and its vibration gains a new high-pitched pattern. The model flags it, and the sensor sends one short alert over its low-power radio to the maintenance team. The bearing is replaced at the next planned stop, before it fails. This is predictive maintenance, and such add-on sensors are described on predictive maintenance sensors.
TinyML chips have other uses. In earphones and small voice devices, one listens for a wake word and wakes the main processor only when it hears it. In a shop, a camera can count people without sending a picture anywhere. Some AI hardware devices use such a chip to listen all the time at very low power.
How does this differ from the AI accelerators in data centres? Those chips use hundreds of watts each to train and run very large models. A TinyML chip runs one small model for one narrow task, using a millionth of that power. Several makers of these chips do their design work in Bengaluru and Hyderabad.