Battery life comes out shorter than expected because the estimate divides rated capacity by average current, and both numbers are usually wrong in the same direction. The cell delivers less than its rating under your load, because pulse currents, rising internal resistance and cold pull the voltage below your circuit’s cutoff while charge remains. Meanwhile the field average is higher than the bench average, because firmware retries, stays awake after errors, leaves peripherals powered and downloads updates. You close the gap by derating capacity for your cutoff, pulses and temperature, measuring a full cycle on production hardware, adding rare events, and logging battery voltage and wake time from the field.
This article collects the causes a firmware team can act on, with numbers from cell makers’ manuals and chip vendors’ documentation, checked in November 2026. It starts from the default example in our battery life calculator: a fictional soil probe on two AA lithium cells, averaging 24.0 µA and estimated at 12.1 years.
Why the estimate and the field disagree
The estimate is honest arithmetic on four assumptions, and each one fails in a predictable way. Capacity divided by current assumes the cell gives its rated capacity at your load, that your bench average is your field average, that current is steady rather than pulsed, and that nothing else drains the cell. Field units break all four.
| Assumption inside the estimate | What breaks it in the field | Where to look |
|---|---|---|
| The rated capacity is available to you | The rating uses a gentle load and a low cutoff voltage; your device stops at a higher voltage, at higher pulse currents | Cell maker’s curves at your load and cutoff |
| The bench average is the field average | Retries on weak links, radios or sensors left on after errors, longer wake-ups, update downloads | Field counters for wake time, retries and resets |
| Current is steady | Radio and sensor pulses drop the cell voltage by current times internal resistance, which grows as the cell drains | Loaded voltage during the largest pulse |
| Nothing else drains the cell | Self-discharge, storage before deployment, cold that temporarily hides capacity | Cell maker’s shelf-life and temperature data, warehouse time |
The calculator’s own note says it plainly: “Real cells lose capacity in the cold and at high pulse currents, so measure on hardware before you promise a number.” The rest of this article explains each row and what to do about it.
Rated capacity is measured under conditions your device does not meet
A cell’s milliamp-hour rating describes one load and one cutoff voltage, and your product rarely matches either. Energizer’s lithium coin cell handbook says coin cell capacity “is influenced by several factors, including drain rate, cutoff voltage, and operating temperature”, and that cells “deliver their best efficiency at lower drain rates”. TI’s white paper on coin cells notes that rated capacity “is typically given for a low, steady state current ranging from a few hundred uA to a few mA” and is “normally measured until voltage drops below 2.0V”.
The cutoff is where many estimates go wrong first. Energizer warns that “products with a high cutoff voltage (greater than 2.0 V) may leave significant capacity unused when the device stops functioning”. Your effective cutoff is the highest of several thresholds:
- The microcontroller’s brown-out reset level, as configured.
- The minimum supply voltage of the radio, the sensor or the external flash.
- The input range of your regulator, if you have one.
- Any voltage below which your firmware refuses to write flash or start an update.
The same logic applies to AA lithium cells. Energizer’s lithium iron disulfide handbook shows an AA example “being discharged at 100 mA to a 0.9 volt cutoff”, and says capacity “will vary with the drain rate and the cutoff voltage”, though less than for alkaline cells. Two cells in series double the voltage, not the milliamp-hours. If your circuit stops at 2.0 V on two cells, each cell stops at 1.0 V, not the 0.9 V in the example curve.
What to do. Write down your device’s real cutoff voltage, find the cell maker’s discharge curve closest to your load and temperature, and set the usable share in your estimate from that curve rather than from habit. The calculator default uses 85%; that is an assumption to replace with your own number.
Coin cells and AA lithium cells fail differently
The two cell types most battery-powered sensors use behave differently under firmware loads, and the estimate treats them the same. A CR2032 is a 3 V lithium manganese dioxide cell with high internal resistance that rises as it drains. An AA lithium iron disulfide cell is a 1.5 V cell built, in Energizer’s words, as a “jellyroll” of long, thin electrodes, and its handbook says its internal resistance “changes very little with age, discharge rate, depth of discharge and temperatures of -20 to 40°C”.
| Property | CR2032 lithium coin cell | AA lithium iron disulfide cell |
|---|---|---|
| Nominal voltage | 3 V | 1.5 V (fresh open-circuit 1.79 to 1.83 V) |
| Internal resistance | Starts near 10 ohms and rises during discharge | Changes very little with age, depth of discharge and temperature from -20 to 40°C |
| What limits life under pulses | Voltage drop during radio peaks, worst near end of life and in the cold | Usually the average current; pulses matter much less |
| Self-discharge in storage | About 1% of capacity per year | About 95% of capacity retained after 20+ years |
| Firmware priority | Keep peaks low and separated; measure loaded voltage | Keep the average low; watch for long awake periods |
The voltage, resistance and storage rows come from Energizer’s two handbooks linked above; the last two rows are our engineering reading.
For example, suppose a wearable on a 220 mAh CR2032 averages 20 µA: the arithmetic says about 1.25 years, but if the cells your customers buy deliver half their rating under your pulses, as TI found for some vendors, the same firmware lasts about seven and a half months. The firmware did not change; the cell’s usable share did.
Pulse loads, internal resistance and the brown-out that ends life early
A cell’s voltage under load is its open-circuit voltage minus the current times its internal resistance. Radios draw short pulses of tens of milliamps, and coin cells have high internal resistance that rises as they discharge. Sooner or later a pulse drags the voltage below your cutoff, the device resets, and the product is dead while the cell still holds charge.
Energizer gives the starting point: “the starting IR of a 2032 battery is near 10 ohms”, and “the battery IR will typically increase during discharge”. The same handbook adds a detail that matters for very low-power designs: “very low uA drain rates can cause a significant rise” in internal resistance early in the discharge. A device that sleeps for hours at a few microamps and then transmits can meet a higher resistance than its total charge used would suggest.
The handbook’s pulse example makes the consequence explicit: a pulsed load “would meet a 2 volt cutoff much sooner than the average drain” would suggest. The Things Network described the same failure for LoRaWAN devices: a transmit peak “would lead to a voltage drop below the IoT device cutoff voltage, creating a brownout and setting the limit of the device battery lifetime”.
What TI measured on CR2032 cells
TI’s white paper SWRA349 is the most useful public data on this problem. It tested CR2032 cells from six vendors, at least 12 cells per vendor, with a load shaped like a Bluetooth LE event: 8 mA for 2 ms, a 1 ms radio peak of 30 mA or 15 mA, 8 mA for 2 ms, then 0.1 mA sleep for 200 ms. A cell’s life ended when the voltage dropped below 2.0 V. The findings:
- “The voltage drop caused by the battery’s internal resistance (IR) during peak load is limiting the effective capacity.”
- Some vendors, “including some branded name vendors, only achieve app. 50% of rated capacity” at either peak current.
- Raising the peak from 15 mA to 30 mA cost only 9% of effective capacity on average, because internal resistance climbs steeply near the end of life. TI put the internal resistance limit at about 30 ohms for 30 mA peaks and about 60 ohms for 15 mA peaks.
- A capacitor in parallel with the cell gave a “solid >40% improvement” with the poorest cells, and 5% and 13% with the best.
TI’s conclusion is worth quoting because it corrects a common rule of thumb: “Bringing the average current down is therefore the most important factor when maximizing battery life of CR2032 coin cells in low power RF applications.” The often-quoted 15 mA limit is not a cliff; cell quality and average current matter more.
Cell-to-cell and brand-to-brand spread
The Things Network tested CR2032 and CR2450 cells from seven manufacturers with a one-second, roughly 40 mA burst and found that “coin cell performance differs a lot between manufacturers, especially at higher currents”. It used one cell per size and maker, which it flags as a limit, and its advice follows from that: “test many”, and do not rely on users to buy good replacement cells. If your product takes user-replaceable coin cells, your estimate has to survive the worst cell a customer can buy.
What firmware and hardware can do about pulses
- Add a reservoir capacitor. TI sized about 100 µF for its harsh test profile (its formula gives 87 µF for a 1 kilohm worst-case internal resistance) and notes a real Bluetooth LE application could use a capacitor “reduced by a factor of 2-5”. The cell recharges the capacitor during sleep, so check that the recharge time fits inside your sleep interval, as TI’s appendix does.
- Never stack loads. Do not power a sensor, write flash and transmit at the same moment. Sequence them so the peaks do not add.
- Use the lowest transmit power that keeps the link. A lower peak raises the internal resistance the cell can tolerate.
- Measure loaded voltage, not idle voltage. Sample the supply during or right after the largest pulse. An idle reading looks healthy until the day the device resets.
- Set the brown-out threshold deliberately. Choose it from the minimum voltage your components actually need, then design the pulse budget around it.
- Account for inrush. Energizer warns that a device’s initial inrush current can be several times higher than its general in-use load, which causes a larger voltage drop than expected. Sensor rails and radios powering up are typical sources. Soft-start the rails firmware controls, and power them up one at a time.
Test real cells with your real profile
Arithmetic tells you where the risk is; a discharge test on real cells tells you whether it bites. The Things Network’s coin cell test shows a practical method: record your device’s current profile, then use a battery profiling setup to drain real cells with it, speeding the test by shortening the low-current periods. It sized its run for a 30-day discharge and stopped either when the cell’s open-circuit voltage reached 2.0 V or when the loaded voltage fell below 0.6 V.
Two cautions apply to accelerated tests. Compressing the sleep time raises the average current, and Energizer notes that internal resistance behaves differently at very low microamp drains, so an accelerated test can be kinder than years of real sleep. And one cell per brand proves little: TI used at least 12 per vendor because “the battery-to-battery variation can be quite large”. Test several cells from more than one batch, and include the brand your customers are most likely to buy as a replacement.
Cold, heat and time drain capacity the estimate ignores
Temperature and time take capacity without any current from your firmware. Energizer’s coin cell handbook says “colder environments” reduce discharge efficiency while “warmer conditions” accelerate self-discharge, and gives the example of “a wireless garage door sensor” that “could stop functioning in the cold of winter due to an excessive voltage drop”. It also notes the cold loss is reversible: “Once the battery warms back to normal operating conditions, voltage and capacity recover.” A device that dies every January and revives in March is showing you a pulse-voltage problem, not an empty cell.
TI adds that “at low temperatures it probably becomes even more important to use a capacitor since the internal resistance increases with lowering temperatures”. Energizer describes its lithium iron disulfide AA cells as having “a much lower sensitivity to temperature compared to other chemical systems”. Its handbook gives a recommended operating range of -40°C to +60°C and says they “can deliver approximately full rated capacity at -40°C if they are discharged at 25 mA”. That is a reason to choose them for outdoor sensors, not a reason to skip a cold test of your own pulse profile.
Self-discharge and storage
Energizer says lithium coin cells “will lose approximately 1% of their capacity per year” at normal room temperature and humidity, with a shelf life of 10 to 12 years. For its AA lithium iron disulfide cells, it says they “retain approximately 95% of its capacity after 20+ years of storage”. Two effects follow:
- Passivation. After long storage, coin cells “can develop a perceptible poor conducting passivation layer” that slightly raises internal resistance at first use, until the load thins it. The first pulses after a long shelf life are the hardest, which is exactly when a new unit gets installed.
- Warehouse time with the battery connected. If the device sleeps on its cell between the factory and installation, that time counts. For example, a unit that sits for nine months at 4 µA uses about 26 mAh; if a factory test leaves the radio advertising at a 24 µA average, the same nine months cost about 158 mAh. Only a hardware cut-off, such as a pull tab or a ship-mode switch, makes shelf drain zero.
Firmware behavior the spreadsheet never sees
The biggest gaps usually come from firmware doing more in the field than on the bench. A spreadsheet holds one clean duty cycle; a field unit lives through weak links, interrupted sensors, error paths and updates. Each of these raises the average current, often by more than the sleep current you spent weeks reducing.
The bars are hypothetical, but the mechanisms are common. In the default example the radio is 62.5% of the average current, so anything that lengthens radio time moves the result most. Doubling each radio event from 2.5 ms to 5 ms takes the radio from 15 µA to 30 µA and the estimate from 12.1 to 7.5 years.
| Firmware behavior | What it does to the average | How to catch it |
|---|---|---|
| Retries on a weak link | Each failed exchange repeats the radio time; a distant gateway can double it | Count retries and radio-on time per day on the device |
| Radio left on after an error | A connection or scan that never times out runs at milliamps for hours | Watchdog-style limits on every radio state; log the longest radio-on period |
| Sensor not powered down | A sensor or its regulator stays on between readings and adds its quiescent current to sleep | Measure sleep current on production units with every sensor fitted |
| Long wake-ups | A sensor that waits for warm-up or a slow bus stretches each wake; 150 ms becoming 1.5 s costs 4.5 years here | Log wake duration, not only wake count |
| Debug features left in | Console UARTs, logging and test commands keep clocks running | Compare release and debug builds on the same board |
| Update downloads and retry loops | One update is cheap; an update that fails and restarts every hour is not | Cap download attempts and require a minimum battery voltage before starting |
Updates cost energy too
A firmware update is a long radio session plus flash writes. Our BLE update time calculator puts the default 412 KB image at 36.6 seconds including verify, swap and reboot. For example, at 6 mA that is about 0.06 mAh, negligible a few times a year. If the same update fails verification and restarts every hour, it costs about 1.5 mAh a day, an average of roughly 60 µA, which is more than twice the probe’s entire budget. On cellular devices, image size matters more: our article on delta firmware updates covers when smaller images pay off, and our comparison of LTE-M and NB-IoT covers the cellular radio budget.
Wi-Fi devices: the lab is kinder than the field
Wi-Fi power depends on how often the radio wakes and how long it stays awake, and both get worse outside the lab. Espressif’s ESP-IDF low-power guide for the ESP32 lists average currents while connected as a station: about 29 to 31 mA in modem-sleep, 20 to 23 mA with modem-sleep plus frequency scaling, and 2.19 to 3.34 mA in auto light-sleep, depending on the access point’s DTIM interval. Deep-sleep, which drops the connection, averages 5 µA. The guide notes the figures were “obtained by testing in a shielded box”.
The same guide lists what stretches each wake-up, including clock drift and “time spent processing missed beacon frames”. A shielded box has no crowded channels and no distant access point. A device configured for auto light-sleep on the bench can spend far longer awake per beacon in a real home, so measure it on the networks your customers use.
The table also shows why battery-powered Wi-Fi devices rarely stay connected. For example, suppose a device holds its connection in auto light-sleep at 2.5 mA, inside Espressif’s measured range. On two AA cells with 3,000 mAh usable, that is about 1,200 hours, or 50 days. A product that must last years on Wi-Fi therefore spends most of its time in deep-sleep, which drops the connection, and reconnects to report. That moves the cost into reconnection time. Measure how long a reconnect takes on a busy network, because that becomes your radio event.
RTOS power management only helps if drivers use it
Low-power frameworks do the right thing only when the firmware tells them what it needs. In Zephyr, system power management puts an idle system into a low-power state, but “it is the application’s responsibility to set up a wake-up event”, and not all wake-up sources work in every power mode. Device runtime power management suspends idle devices based on a usage count, where “the device driver is responsible to indicate when it needs the device and when it does not”. A driver that takes a reference and never releases it on an error path keeps its peripheral awake indefinitely. Test the error paths, not only the happy path.
Wake-up sources have the same trap. Zephyr’s documentation lists typical wake-up events such as “a SysTick, RTC, counter, or GPIO” interrupt, and warns that “some wake-up sources may not be usable in all power modes”. If the firmware schedules its next reading on a timer that only runs in a lighter sleep state, the power policy may be unable to choose the deepest state, and the board sleeps at a higher current than the one you measured on a test build. Check which state the device actually enters, on the production build, by measuring it.
Diagnose a fleet that dies early
When field units are already dying, the first question is whether the cells ran out of charge or ran out of voltage. The two point to different fixes: too much charge used is a firmware or duty-cycle problem, while charge left in a cell that cannot hold voltage under your pulses is a load-profile or cell problem.
- Collect dead units and their cells, plus a few still-working units from the same cohort. Record install date, location, firmware version and the last reports each unit sent.
- Let each cell rest, then measure its open-circuit voltage. Energizer notes that after a load, voltage “drops dramatically and then slowly recovers with time”, so measure after the cell has recovered. For AA lithium iron disulfide cells, Energizer says a good cell “will generally have an OCV >1.74 volts”, and one below 1.70 V after recovery “is completely discharged”.
- Load the cell with your real pulse profile. A bench supply cannot show this; the cell itself must drive a resistor or an electronic load stepped to your peak current, while you record the voltage.
- Read the reset log, if the firmware keeps one. Repeated brown-out resets before death point to the pulse problem; a clean, gradual fade points to average current.
- Compare cohorts. Group units by firmware version, install month, distance to the gateway and site temperature. A cluster of early deaths at the edge of range is a retry problem; a cluster after one release is a regression; a cluster in one cold month is a pulse and temperature problem.
- Reproduce on the bench with the field firmware, the worst-case link and, if you can, the cold. Then fix and verify on a small staged wave before shipping to everyone.
| What you find | Likely cause | Fix to try first |
|---|---|---|
| Cell open-circuit voltage healthy, loaded voltage collapses at your peak | Internal resistance too high for your pulses, often near end of life or in the cold | Reservoir capacitor, lower or separated peaks, a better cell |
| Cell fully discharged well before the estimate | Field average current higher than the bench average | Field counters for radio time, retries and awake time; find the outlier state |
| Deaths cluster after one firmware release | A regression: a peripheral not suspended, a longer wake, a retry loop | Compare current traces of the two versions on the same board |
| Deaths cluster at the edge of range | Retries and long radio sessions | Retry limits, backoff, adaptive reporting, a closer gateway |
| Units stop in winter and recover in spring | Cold raising internal resistance until pulses brown out | Capacitor, cold-rated cells, a lower brown-out threshold if components allow |
| New units fail sooner than old ones | Shelf time with the battery connected, or passivation in long-stored cells | Ship mode or a pull tab; check storage history |
Build a derated estimate you can defend
A defensible estimate starts from measurements on production hardware and derates for every effect the arithmetic leaves out. It takes a day or two, and it turns “12 years” into a number you can put in a requirement.
- Measure one full cycle on production units. Capture a current trace of a complete wake, sense, transmit and sleep cycle on several boards with production firmware, with the debug probe disconnected. Our article on measuring sleep current covers the setup and the usual errors.
- Find your real cutoff voltage. Take the highest of the brown-out level, component minimums and your firmware’s own thresholds.
- Set the usable share from the cell maker’s curves at your load, your cutoff and your coldest operating temperature.
- Check the pulse budget. Using the cell’s end-of-life internal resistance, confirm your largest pulse keeps the voltage above the cutoff, or add a capacitor.
- Add field radio time. Increase radio events for the retry rate you expect at the edge of range, or measure it at the edge of range.
- Average rare events over a month. Updates, reconnects after a gateway outage and error recovery all go in as extra radio or wake time.
- Count the shelf. Add storage time with the battery connected and the cell’s self-discharge.
- Compare with the requirement and keep the margin you will need for the unknowns.
| Input | Bench habit | What to use instead |
|---|---|---|
| Usable capacity | The rating times a round percentage | The cell maker’s curve at your load, cutoff and coldest temperature |
| Sleep current | A datasheet figure or a development board | Several production units, production firmware, debug probe disconnected |
| Radio time per event | One clean exchange next to the gateway | The median at the edge of range, including retries |
| Wake duration | The happy path | The 95th percentile from field logs, once you have them |
| Rare events | Left out | Updates, reconnects and error recovery averaged over a month |
| Shelf time | Left out | Months between factory and installation, at the shelf current |
Example (fictional). Loamwick’s LW-1 probe has a requirement to run 3 years on two AA lithium cells, and the calculator’s default estimate is 12.1 years at 24.0 µA. For winter on the Eastern Plains and a 2.0 V cutoff, the team lowers the usable share from 85% to 70%. Probes at the far edge of the trial plots retry often enough that radio events average 5 ms instead of 2.5 ms. Four BLE updates a year add about 0.03 µA, which rounds away. The derated average is 39.0 µA and the estimate falls to about 6.1 years: still twice the requirement. The team’s conclusion is that the requirement survives realistic derating, and that the real risk is a firmware regression such as a radio stuck on for an hour a day, which would take the estimate to about a year.
Make the last weeks of the battery behave well
Firmware decides what happens at the end of a cell’s life, and a bad ending wastes the last of the charge or corrupts the device. Energizer notes that lithium coin cells “maintain a relatively flat discharge curve until near end of life” and asks device makers to “account for this quick drop in voltage at end of life to avoid abrupt shutdowns”.
Three behaviors are worth designing on purpose:
- Stop the brown-out reboot loop. A device that browns out during a transmit, reboots, rejoins the network and transmits again can loop until the cell is flat, and each reboot is a fresh chance to interrupt a flash write. On boot, check the reset reason; after repeated brown-out resets, back off for hours before trying the radio again.
- Keep a reserve for the last message. Pick a loaded-voltage threshold above your cutoff, and when a device crosses it, send one “battery low” report, stop optional work such as updates and extra diagnostics, and lengthen the reporting interval.
- Refuse to start an update on a weak cell. An update is the longest radio session and the most flash writing the device will do. Require a minimum loaded voltage before downloading, and again before the swap.
Voltage alone is a poor fuel gauge
Because the voltage curve is flat until near the end, a “battery percentage” computed from voltage reads full for most of the cell’s life and then drops quickly. For a better estimate, count charge in firmware: multiply the time spent in each state by that state’s measured current and add it up. This software coulomb count is only as good as your current measurements, but it moves steadily, and it tells you which state used the charge. Report it alongside loaded voltage and correct it when the cell is replaced.
Close the loop with field data
An estimate is a hypothesis until field units report on it. The cheapest way to find out why batteries die early is to have devices tell you, before they die, how they spend their charge. Loamwick’s v1.1.0 release in our example project adds battery voltage to every report for exactly this reason.
Report these from every device, in every message or once a day:
- Loaded battery voltage, sampled during or right after the largest radio pulse, alongside the idle voltage. The gap between them tracks internal resistance as the cell ages.
- Reset reasons, especially brown-out resets. A rise in brown-outs across a cohort is the earliest sign of the pulse problem.
- Radio-on time and retry count per day. These show which units sit at the edge of range.
- Total awake time per day, and the longest single awake period.
- Firmware version, so you can compare cohorts before and after a release.
- Temperature, if the device has a sensor, so you can separate cold dips from real depletion.
Then compare releases. A release that raises daily awake time by a few percent across the first wave of a staged rollout is a battery regression, even if every functional test passed. Our firmware release checklist includes “Battery and current draw checked against the last release” for this reason; treat a change in daily radio time like any other failed test.
Finally, feed the field numbers back into the estimate. Replace the bench radio time with the field median, the usable share with what returned cells actually delivered, and rerun the calculation. After a season of reports, the estimate stops being a promise and becomes a forecast based on your own fleet.
What to do this week
Start with the two numbers most estimates get wrong: the cutoff voltage and the field radio time. Then measure, derate and instrument.
- Write down your real cutoff voltage and set the usable share from the cell maker’s curve at your load and coldest temperature.
- Measure a full cycle on production units with production firmware, not a development board.
- Check your largest pulse against end-of-life internal resistance, and add a reservoir capacitor if it fails.
- Put a time limit on every radio state and cap update retries; require a minimum loaded voltage before an update starts.
- Add loaded voltage, reset reason, radio-on time and awake time to your telemetry.
- Re-run the battery life calculator with the derated inputs and keep the result next to the requirement.
Fumbar writes embedded firmware and update systems for hardware startups. Our low-power design work covers power profiling, sleep states and battery budgets on your hardware, and environmental sensors are among the products on our climate and energy page. If you want a second read of your power budget alongside your schematics and code, a firmware review is a short, fixed-scope first engagement that ends in a written plan.
