Power Consumption (Edge)

A battery-powered device has a finite energy supply, and every model prediction draws from it. Power consumption at the edge describes how much electrical power an AI workload uses while it runs—and, just as importantly, how that use changes battery life, heat, and device reliability.

Power versus energy
Power is the rate at which the device draws energy, measured in watts (W) or milliwatts (mW). Energy is the total amount used over time, measured in joules (J) or milliwatt-hours (mWh). A model that briefly draws 1 W can use less energy than one that draws 200 mW continuously for hours. For edge AI, both views matter:

Peak power determines whether the battery, voltage regulator, or thermal design can support a burst of computation.
Average power determines battery life during normal use.
Energy per inference shows the cost of one prediction, useful for event-driven sensors and cameras.

What consumes the power
Running the neural network is only part of the bill. A smart camera also spends energy capturing frames, moving image data through memory, waking the CPU or NPU, and transmitting results. Memory access can cost more energy than arithmetic, so a compact model with poor memory behavior can still drain a battery quickly. On a Cortex-M microcontroller, keeping a wake-word model in low-power memory and processing small audio windows helps avoid expensive wake-ups. On a phone, TensorFlow Lite, Core ML, or ONNX Runtime can route supported operations to an NPU, which usually performs inference more efficiently than the general CPU.

Why measurement matters
Ignoring power can produce a model that looks fast in a short benchmark but fails in deployment: the device heats up, throttles its clock speed, or needs charging far too frequently. Engineers measure power with external power monitors or board-level sensors while testing realistic duty cycles—such as a wake-word detector listening continuously, then running a larger speech model only after activation. Choices such as quantization, lower input resolution, frame skipping, batching, and accelerator use reduce energy, but must still preserve the accuracy and responsiveness the product requires.

Power consumption (edge) is the electrical energy an edge device uses while running an AI model, including compute, memory access, and supporting hardware activity. It is measured as power over time or energy per inference. Low consumption extends battery life, limits heat and throttling, and enables reliable on-device inference on phones, sensors, and microcontrollers with fixed energy budgets.

Think of power consumption as an AI device’s appetite for electricity. A phone camera that recognizes faces, a smart doorbell that spots packages, or a fitness watch that tracks movement all need energy each time their AI is working.

For Edge AI, this matters because the AI runs on the device itself, often using a battery or a small fixed power supply. An AI feature that uses too much power can drain a phone faster, make a wearable need frequent charging, or cause a sensor to run out of energy in the field. Lower power consumption means the device can stay useful longer without sacrificing its smart features.