Edge AI Duty-Cycle Lab: TinyML Battery Budget

Power Mode

Duty-cycled (sleep between windows)

Sensor & Model

Sampling rate 50 Hz

Wake interval T 500 ms

Detection threshold 0.60

TinyML Model Size

Live Readouts

Inference time

— ms

Duty cycle

— %

Avg power draw

— mW

Est. battery life

— h

Detections so far

0 detected · 0 sent

How it works

An always-on IoT classifier and a duty-cycled one run the identical TinyML model, but spend very different amounts of time in a high-power active state versus a microamp-scale deep sleep. The node samples a window of the sensor signal, runs a fixed-cost neural-network inference over it, and only then reports a detection — the energy budget is a straight sum of the two states:

t_infer = FLOPs / (f_clk · IPC)              — compute time per inference
t_active = t_window + t_infer                — sampling + inference burst
duty     = t_active / T                      — fraction of the wake period spent awake
P_avg    = duty·P_active + (1-duty)·P_sleep  — time-weighted average power
I_avg    = P_avg / V_batt                    — average current draw
life     = Q_batt / I_avg                    — hours until the cell is empty

Sampling rate — how many raw samples fill the classifier's input window; more samples per window means a longer active burst before inference can start.
Wake interval T — the period between wake-ups in duty-cycled mode. A short T means the node barely sleeps; a long T saves power but raises detection latency.
Model size — Tiny/Small/Medium/Large scale the model's FLOP count (and its simulated accuracy), which scales t_infer directly — a bigger model is a longer, hungrier active burst on the same microcontroller clock.
Always-on mode — disables sleep entirely: the radio and MCU core stay powered between windows, so P_avg collapses to P_active regardless of T, which is why always-on nodes need mains power or die in hours instead of months.
Detection threshold — the classifier score an inference must clear before the node counts it as a real event and transmits a packet (the flying dots to the gateway icosahedron); raising it suppresses false triggers but can miss a genuine anomaly the model scored just under the line.

This is the same trade-off real TinyML tooling like Edge Impulse exposes as an "EON tuner" — the same trained model can be deployed at several quantization/window settings, and the deployment that wins is whichever meets the accuracy target at the lowest measured microamp-hours, not necessarily the most accurate one.

idle

MCU / sensor node Detection → packet to gateway Battery reserve

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.