← R&D Log
2026-11-03EDGE6 min read

SWaP-C is a procurement line item

Size, weight, power, and cost decide whether an edge AI system can be fielded at all. Treating power as a model-accuracy problem is how field deployments fail.


Most edge AI projects are scoped by accuracy and sized by hardware. The field deployment then fails on power. A sensor node that runs for three days on the bench runs for three hours on a battery in the field, and no amount of extra accuracy fixes that. SWaP-C (size, weight, power, and cost) is the constraint that decides whether the system can be bought, shipped, and kept running.

Power is the budget you can't refill

On a remote node, the battery is fixed and recharging may be unavailable for weeks. Every milliwatt of idle draw is a day of runtime lost over a deployment. That turns the usual engineering trade-offs around:

  • A model that is slightly less accurate but spends most of its time asleep can be the better system.
  • Continuous high-accuracy inference is often the wrong default for an always-on sensor.
  • The wake-up path matters as much as the classifier.

Cascade the work

The approach that has worked for us is a two-stage cascade. A very small detector runs continuously. It is tuned to fire on anything that might be a target, and it is cheap enough to ignore most of the environment. Only when it fires does the heavier classifier run.

On a multi-month autonomous remote sensor node, this structure reduced continuous power draw by over 75%. The gain comes from the duty cycle, not from a faster classifier.

The trade-off is recall. The gate must be permissive enough that real targets are not missed, and the classifier must be good enough that false wake-ups do not drain the battery. Those two numbers are set together, against the battery budget, not chosen separately.

Quantization and runtime are procurement questions too

Compression choices affect cost and supply. INT8 or FP16 quantization can move a model onto cheaper silicon, and a compiled runtime (ONNX, TensorRT, or C++ for a specific SoC) can keep the bill of materials down. Each choice narrows the set of boards you can buy, so the runtime decision belongs in the purchasing conversation, not only the engineering one.

What to put in the specification

If you are writing a requirement for an edge AI system, state the constraints in terms a buyer can verify:

  1. Required runtime on battery, at a stated duty cycle and temperature range.
  2. Maximum idle draw, measured, not estimated.
  3. Detection performance at the operating point, with the false-wake rate stated.
  4. Whether updates can be pushed to deployed units, and over what link.

A supplier who answers these with measurements is showing you the system. One who answers with accuracy alone is showing you the model.


Clean-room write-up — public / synthetic data only.