A thermal throttle is not Low Power Mode
When a beta feels fine on the bench and "janky outdoors," teams often open one ticket labeled performance under load. Heat and Low Power Mode both make the OS throttle work, so the instinct is to treat them as the same constraint. They are not. thermalState climbs because the chassis is hot and the system is protecting itself right now. Low Power Mode is a power policy the user or system turned on to stretch the remaining charge. On Endeo long capture sessions and RuleCaddie fairway use we have seen both — sometimes together, often alone — and the right degradation for one is the wrong cut for the other. Device testing that only flips one condition, or that folds both into a single shouldReducePerformance bit, ships a guess.
Two signals, two questions
ProcessInfo.thermalState answers how hard the silicon is already being limited by temperature. The stages — nominal, fair, serious, critical — are the OS negotiating survival. From serious onward, CPU and GPU headroom shrink whether your app asked for that or not. Keeping heavy ink recognition, speculative OCR, or decorative animation running there makes the heat worse and the frame time fall off a cliff.
isLowPowerModeEnabled answers whether the device is under a power-saving policy. It can be true on a cool phone with twenty percent battery, and false on a hot tablet plugged into a charger. The remedies that fit are different: defer background refresh, stretch sync intervals, skip prefetch the app started on its own — not necessarily gut the same GPU path you cut under critical thermal.
Battery percentage alone is not Low Power Mode either. A low battery without LPM still has the normal scheduling budget. LPM without heat still has thermal headroom. Mixing "low juice" anecdotes into a thermal ticket, or calling every stutter a battery bug, erases which ProcessInfo value was actually true when the session failed.
What goes wrong when you collapse them
A single constrained flag that ORs heat and LPM tends to apply the heavier cut everywhere. You stop prefetch under fair thermal when you only needed to drop decoration, or you freeze a user-initiated sync because LPM was on and the tier table treated that like critical. Users feel a broken feature; the log says you were being responsible.
Lab coverage collapses the same way. Someone runs Xcode Simulate Thermal State through serious, watches the degraded UI, and checks the box for power testing. Low Power Mode was never toggled. Or they enable LPM in Settings, skip a sustained warm session, and never see the thermal climb that only appears after forty minutes of capture on glass.
Field telemetry gets noisier too. MetricKit and Organizer aggregates that do not slice by whether the session was thermally elevated or power-constrained blend two failure modes into one hitch rate. WWDC-era state reporting exists so you can tag meaningful app configuration — including power-constrained versus nominal — instead of one blended average that cannot tell you which drawer to open.
How we induce each condition in the lab
Thermal: attach a physical device, use Xcode's Simulate Thermal State (Device Conditions) and step fair → serious → critical while exercising the primary loop. Confirm the app reads thermalStateDidChangeNotification, cuts the right discretionary work at each stage, and keeps user-initiated actions alive. Then run at least one unforced sustained session on mid-tier hardware — unplugged, bright screen — because a forced serious on a cool chassis is not the full cascade of memory pressure and frame pacing you get when the device cooked itself there.
Low Power Mode: toggle it from Settings on device (or the simulator Features menu for wiring checks). Confirm NSProcessInfoPowerStateDidChange fires, background tasks and prefetch back off, and explicit user taps still complete. Do not treat a simulator LPM toggle as proof of energy behavior on sealed hardware; treat it as a state-machine check, then re-run the same policy on glass.
When both can be true in the field — hot phone, LPM on — test the combination as a third row, not as a substitute for the singles. Record which signal drove each cut. If the note only says "tested under load," you did not test either condition.
Different backs-offs, same discipline
Under rising thermalState, make what is already running lighter: drop decorative animation, reduce recognition fidelity, pause speculative OCR, lower non-essential frame work. The goal is to stop feeding the heat while the core capture or navigation path still works.
Under Low Power Mode, do less work you were about to start: skip BGTask refresh you scheduled yourselves, stretch polling, defer prefetch until the user is about to need it. Do not cancel the sync the user just requested because the policy bit flipped — that is the inconvenience you were trying to avoid.
Centralize the mapping from ProcessInfo into a small quality tier if you want one knob for features — but keep the inputs separate in logs and in the test matrix. A tier is a convenience for UI code. It is not permission to stop naming which OS signal you forced in the device note.
What we check before calling power testing done
The release note has three rows filled: forced thermal steps with observed cuts, LPM on with observed deferrals, and at least one sustained warm session on representative hardware with thermalState logged on an interval. Blank rows mean the claim is incomplete.
No ticket labeled "janky when battery low" closes as thermal without reading thermalState. No thermal ticket closes as "fixed by LPM handling" without a Device Conditions pass. User-initiated paths still succeed under both constraints.
If those checks fail, you tested the cool, full-charge demo — not the phone in a pocket or the iPad on a sunlit table. Heat and Low Power Mode are adjacent pressures. Keep the pipes labeled so neither masquerades as the other.
Questions
- Is Low Power Mode the same as a low battery percentage?
- No. LPM is a ProcessInfo policy bit. Battery can be low without LPM, and LPM can be on without heat. Test and log the policy, not a percentage anecdote.
- Does Simulate Thermal State replace a long session on hardware?
- No. Forced states prove your degraded paths are wired. Sustained unplugged runs prove the real thermal climb and energy budget. You need both.
- Can we use one shouldReducePerformance flag?
- You can map both signals into a tier for feature code, but keep the inputs separate in telemetry and in the test matrix so you know which condition you actually covered.
Sources
- Quietly dialing back work under thermal and Low Power Mode — Separate ProcessInfo.thermalState and isLowPowerModeEnabled; Xcode Simulate Thermal State vs LPM toggle.
- WWDC26 Power and Performance lab notes — Listen for thermalState; use condition inducers and Low Power Mode as distinct stress tests; MetricKit state reporting.
- Related: A cold start is not a session