What does it mean to "balance communication and computation" for energy optimisation on an IoT device, and how do you decide which way to go?
Transmitting a byte usually costs far more energy than computing on it, so the device should process locally until the cost of the extra computation exceeds the transmission it saves.
Both the radio and the processor draw from one battery, which turns a design question into an arithmetic one. The radio is the expensive side: activating a transceiver, acquiring the link and sending a packet costs orders of magnitude more energy per byte than the processor costs per operation. So the default move is to compute locally and transmit less:
- Send an average rather than every sample.
- Send only when a reading crosses a threshold, rather than on a fixed schedule.
- Compress, aggregate, or run a small model on-device and send the conclusion instead of the raw data — the "local intelligence at the edge of the cloud" idea, extending now to on-device AI/ML inference and embedded AI accelerators.
The balance tips the other way when the local processing is heavy relative to the data it saves. Filtering a thousand samples down to one average is obviously worth it; running an expensive transform whose output is nearly as large as its input is not. Nor can you always choose freely — if the raw data is needed for later re-analysis, discarding it locally is a functional decision, not just an energy one.
The same logic runs alongside low-power sensing (choosing sensors and sampling rates that do not dominate the budget) and efficient energy management, storage and battery use, with energy harvesting where the deployment allows it.
Tip: the rule of thumb is "a radio byte costs more than a CPU cycle." Design as if bytes are billed and cycles are cheap — and then check the arithmetic for your specific hardware, because the ratio varies by radio.
Go deeper:
Wikipedia — Sensor node — states the asymmetry this card turns on: more energy goes to data communication than to any other process on the node.