Sub-mW Inference

Imagine a tiny sensor that listens for a wake word all day from a coin-cell battery. It cannot afford the power draw of a phone or cloud connection, so its machine-learning work must be nearly invisible on the energy meter. That is the goal of sub-mW inference: running a model while using less than one milliwatt of power.

What the measurement really means

A milliwatt is one thousandth of a watt. In practice, “sub-mW” usually describes the average power used for an inference workload over time, not necessarily the brief peak while the processor is active. A microcontroller might wake, process a short audio or sensor window in a few milliseconds, then return to deep sleep. The energy used per decision, multiplied by how frequently decisions are made, determines average power. A credible measurement also states what is included: the model runtime alone, or microphone, sensor, memory, regulator, and processor together.

How devices reach that budget

Sub-mW designs combine a very small workload with careful hardware scheduling:

Quantized models, commonly 8-bit integers, avoid costly floating-point math and reduce memory traffic.
Duty cycling keeps the MCU asleep between sensor windows instead of running continuously.
Feature extraction converts raw signals into compact inputs, such as audio features for keyword spotting.
Optimized kernels, such as CMSIS-NN used with TensorFlow Lite for Microcontrollers on Arm Cortex-M chips, perform the required operations efficiently.

Why it matters on TinyML hardware

This power level unlocks months or years of unattended battery operation for vibration monitors, leak detectors, wearables, and always-listening wake-word devices. Ignoring it can turn a model that fits in flash memory into a product that drains its battery in days, heats up, or misses real-time deadlines. Sub-mW inference is therefore not just a smaller model benchmark: it is a system-level promise that useful local intelligence can remain awake without consuming the device’s energy budget.

Sub-mW inference is machine-learning model execution whose power consumption stays below one milliwatt, typically on ultra-low-power microcontrollers or sensor processors. It combines compact models, efficient operators, and aggressive duty cycling to classify local sensor data within extremely tight energy budgets. This capability enables always-on functions such as wake-word, vibration, or anomaly detection on coin-cell and energy-harvesting devices, where higher-power inference would rapidly exhaust the available energy.

Think of a smoke alarm that must stay awake for years on a tiny battery. It cannot afford to constantly send data to the internet or use much electricity. Sub-mW inference means an AI system can make a prediction—such as “I heard glass break” or “this machine is vibrating strangely”—while using less than one milliwatt of power.

That is an extremely small energy budget: less power than many tiny indicator lights use. It matters for sensors placed in fields, wearables, medical monitors, and battery-powered detectors. They can listen or watch for important events for long periods without frequent charging, while keeping decisions on the device itself.