Notes
🔍

Power Constraint

A battery is not just a smaller wall socket. On an edge device, every model prediction draws from a limited energy store, competes with the screen, radios, and sensors, and can create heat that slows the processor down. A power constraint is the practical limit on how much electrical power an AI system can draw while still doing its job safely and for long enough.

Power versus energy
Power, measured in watts (W), is the rate at which a device uses electricity at a particular moment. Energy, measured in watt-hours or joules, is the total amount used over time. Both matter, but they cause different problems:

  • High power during inference can overheat a phone or tiny board, triggering thermal throttling and slower predictions.
  • Steady energy use determines battery life. A wake-word model drawing a small amount continuously can still drain a wearable over a day.
  • Short, high-power bursts can exceed what a weak battery, USB supply, or sensor-node regulator can provide reliably.

What determines the budget
A model consumes power when the CPU, GPU, or NPU performs arithmetic and moves data between memory and compute units. Moving large tensors can cost as much as, or more than, the calculations themselves. Camera capture, wireless transmission, and keeping memory awake add further load. Engineers therefore measure not just accuracy and latency, but watts, joules per inference, and battery life under realistic duty cycles. A smart camera that detects people once per second has a very different budget from one analyzing 30 video frames per second.

Designing within it
Meeting a power constraint shapes deployment choices: use a smaller or quantized model, run inference less frequently, wake a larger model only after a cheap sensor trigger, or target a hardware accelerator. TensorFlow Lite can use delegates for device-specific accelerators; on microcontrollers, TensorFlow Lite for Micro runs compact integer models within extremely tight budgets. Ignoring the constraint produces a model that looks excellent in a benchmark but overheats, throttles, misses real-time deadlines, or leaves a battery-powered product unusable.

Power constraint is the fixed limit on energy consumption and instantaneous power an edge device can supply to AI computation while meeting battery-life, thermal, and hardware reliability requirements. It constrains model size, inference frequency, memory access, and processor choice. Managing it is essential for sustained on-device inference: exceeding the budget drains batteries, causes thermal throttling, or prevents deployment on small sensor nodes and microcontrollers.

Think of a phone or small sensor like a camper with a limited water bottle: every task has to fit within what it can carry. A power constraint is that limit on how much energy a device can use.

For AI running directly on a device, this matters because recognizing faces, listening for a wake word, or analyzing a camera feed can drain a battery. Using too much power also creates heat, which may slow the device down. So on-device AI must be useful without constantly charging, overheating, or quickly emptying a battery. A smartwatch that detects a fall all day is only helpful if it still has power when the fall happens.