A PoE switch almost never dies of the thing the ticket says. Not the port. Not the firmware. Not the camera. In the units I have opened on my bench after a "sudden" failure, the pattern is dull and repeatable: the electrolytic capacitors near the 54 V rail have domed tops, the date code puts the unit between 30 and 42 months old, and the cabinet it lived in had no moving air.
Two weeks ago a contractor sent me photos from a rooftop cabinet in Bao'an — a 24-port PoE switch, nineteen cameras, and a probe reading of 71 °C on the chassis with the door shut. The switch was 34 months old. Four cameras had begun rebooting at roughly the same time every afternoon, which is when the sun lands on that wall. He wanted to know which camera was faulty. None of them were. The switch was not overloaded either — about 60% of its PoE budget. It was cooking, and the parts that cook first are the ones with liquid inside.
What Actually Wears Out Inside a PoE Switch
An aluminium electrolytic is the only common component on that board with an hours rating. Resistors do not wear out. MOSFETs fail from events. A capacitor with a liquid electrolyte fails from a clock, and the clock is temperature.
The electrolyte escapes through the rubber seal at the base of the can. Warmer electrolyte pushes harder against that seal, so the escape rate roughly doubles for every 10 °C. That is the entire mechanism, and it explains why the failure looks arbitrary in the field: nothing breaks, the part just drifts until the circuit stops tolerating the drift.
The datasheet hours deserve a closer read. A part quoted at 2,000 hours at 105 °C is not a part that dies after 83 days. It is a part tested at its maximum rated temperature with rated ripple applied until it reaches an end-of-life criterion — usually capacitance down 20%, ESR up two to three times. Those hours describe the worst case the part is allowed to see.
The field symptom is rarely a dead port. ESR climbs, output ripple climbs with it, and the loop regulating port voltage loses margin it was designed with. Under a load step — infrared LEDs at dusk, a pan-tilt motor starting — the rail dips further than it used to. The PD browns out, drops its signature, and the port re-classifies. Multiply that across a bank of cameras and you get "four cameras reboot every afternoon," which is a power supply symptom wearing a camera costume.
A PoE switch has three capacitor populations ageing at different rates: the primary bulk part at high voltage and moderate ripple, the output capacitors on the 54 V rail carrying the sum of every port, and the small parts on each port's conversion stage seeing the sharpest current edges. All three work every hour of the year. A data-only switch idles at 20–30% load. That is why PoE units in the same building die younger than their non-PoE neighbours, and why nobody believes it when you say so.
The 10 °C Rule, Run Against a Cabinet You Can Actually Find
The arithmetic is old: L = L0 × 2^((T0 − Ta) / 10), where L0 and T0 are the rated hours and temperature, and Ta is the capacitor's core temperature — not the room, not the datasheet ambient. Core temperature is the air around the can plus the part's own ripple heating. In a small radial can, 10–15 °C of self-heating is ordinary.
| Capacitor rating (L0 @ T0) | Core temp in service | Multiplier | Projected hours | Years at 24/7 |
|---|---|---|---|---|
| 85 °C, 2,000 h | 65 °C | 4× | 8,000 | 0.9 |
| 105 °C, 2,000 h | 85 °C | 4× | 8,000 | 0.9 |
| 105 °C, 2,000 h | 75 °C | 8× | 16,000 | 1.8 |
| 105 °C, 2,000 h | 65 °C | 16× | 32,000 | 3.6 |
| 105 °C, 5,000 h | 65 °C | 16× | 80,000 | 9.1 |
| 105 °C, 5,000 h | 55 °C | 32× | 160,000 | 18 (capped) |
Read the fourth row twice. A respectable 105 °C part, run at a 65 °C core, lands at 3.6 years of continuous service. That is where "year three" comes from. Not a defect, not bad luck — a 2,000-hour component doing what its datasheet promised in a box 15 °C warmer than anyone assumed.
The bottom two rows explain why capacitor quality is worth arguing about on a purchase order: same temperature, same circuit, and a 5,000-hour part instead of a 2,000-hour one buys 2.5× the service life for a couple of dollars across the whole unit. Note the cap on that last row — major manufacturers put a ceiling around 15 years on Arrhenius projections, because seal degradation sets a limit no amount of cooling removes. Cooler is better; cooler is not immortal.
How I run this for a real cabinet, in the order I actually do it:
- Ask for the capacitor series and its L0/T0. If the supplier cannot name the series, treat it as 85 °C, 2,000 h — the pessimistic default, and often right.
- Measure cabinet air at mid-height, door closed, hottest hour of the day. Not the room. The box.
- Add 10 °C for the local rise near the PSU heatsink, then 10–15 °C for ripple self-heating. That sum is Ta.
- Apply L = L0 × 2^((T0 − Ta)/10), and round toward pessimism.
- Divide by 8,760 for years of continuous duty. Cameras run around the clock; do not use a working-day figure.
- Cap at 15 years, then compare against the replacement cycle the client actually plans. If the projection is shorter, the fix is thermal, not electrical.
It goes wrong in both directions. A unit with a large extruded heatsink and 105 °C, 5,000-hour caps outlived its fans twice over. And I once mis-called a healthy supply because I measured ambient with the door open, which is a different machine.
PoE Load Never Idles, and That Changes the Clock
Every PoE switch has a heat ledger, and most arguments about cabinet temperature come from reading the wrong line. Here it is for a 24-port unit delivering 300 W in a 28 °C room.
| Where the watts go | Rough magnitude | Stays in the cabinet? |
|---|---|---|
| Delivered to the PDs (cameras, APs) | 300 W | No — dissipated at the device end |
| AC-DC conversion loss at ~88% efficiency | ~41 W | Yes, at the PSU |
| Switch ASIC, PHYs, fans, LEDs | ~15 W | Yes, spread across the board |
| Cable I²R on loaded runs | 0.5–1 W per port | No — spread along the cable |
The first row is misread in both directions. That 300 W does not heat the cabinet; it heats the cameras. But the 41 W of conversion loss does, and it lands in the smallest volume in the building. Fifty-six watts inside a sealed 600 × 400 × 150 mm wall box is a 40 W bulb in a shoebox. In my experience such a box settles 12–18 °C above the room, which is how a 28 °C room becomes a 45 °C intake and a 65 °C core.
The other half is duty cycle. A camera draws the same power at 3 a.m. as at noon — no idle period, no weekend, no sleep state. Thermal cycling is mild, which sounds kind, except it removes the one thing that would otherwise buy time. Meanwhile the fan, the only part in there with bearings, degrades. Airflow drops, the caps age faster, and the last summer of the unit's life is the hottest it ever sees. Failures cluster in July for a reason.
Cold starts are a separate withdrawal from the same account. A switch on a timer that gets power-cycled daily accumulates inrush events that the steady-state numbers do not describe — I covered the millisecond side of that in this piece on inrush current.
MTBF Is Not a Lifetime, and 500,000 Hours Is Not 57 Years
Every datasheet carries a number like 500,000 hours MTBF, and someone divides it by 8,760 to announce 57 years. It cannot answer the question being asked.
MTBF is one over the failure rate, and the rate is summed from component data in a handbook — MIL-HDBK-217F or Telcordia SR-332, at an assumed ambient of 25 °C or 40 °C. Those models describe the flat middle of the bathtub curve: random failures during useful life. Wear-out is excluded by definition. A supply with 500,000 hours MTBF and 2,000-hour capacitors is internally consistent and says nothing about month 34.
The standard also matters. The same design produces very different numbers under MIL-HDBK-217F and Telcordia, so comparing two vendors without knowing the standard, temperature basis and environment category is not a comparison. What I ask for instead is duller and more useful: capacitor series with rated temperature and hours, the derating curve at 50 °C, whether the PSU is internal or a brick, and the fan's rated life. A supplier who can answer those four knows what they built.
What I Measure on a Switch That Is Three Years Old
Six readings decide whether to replace or investigate.
- Cabinet air at mid-height, door closed, 15:00. A logging thermometer left for an hour beats a handheld reading taken at 9 a.m. with the door open.
- Chassis surface at the PSU end, contact probe. Infrared guns read dark plastic badly and shiny metal worse; that display number is often the probe's own reflection.
- 54 V rail at the PSE, at present load and again at half load. A rail that sags noticeably between the two has less regulation margin than it had.
- Ripple on that rail — AC coupled, 20 MHz bandwidth limit, short ground lead. The 100 Hz component is the interesting one. If it has grown since handover, the output capacitors are telling you where they are in their life.
- Far-end port voltage under load, beside the port voltage at the same moment. The difference is the cable; the absolute level is the supply. Separating those two early saves a lot of unnecessary cable replacement.
- Port event counters on a managed switch. Re-classification events, port resets, uptime resets. Three cameras re-classifying every afternoon is a power event, not three coincidences.
The comparison only works with a baseline, which is why I now record these at handover rather than at the first failure. The full set is in the six numbers worth signing off on a PoE installation; cabinet temperature is the one skipped most often, usually because it was a comfortable 30 °C on commissioning day.
What I Specify Now, and Where I Was Wrong
For years my answer was "specify 105 °C capacitors" and I considered the matter closed. Incomplete, as the table shows: a 105 °C, 2,000-hour part at a 75 °C core is 1.8 years. A temperature rating without hours is half a specification, and hours without a measured core temperature is a guess wearing a number.
I also made the expensive version of that mistake once. A client's 16-port unit failed at 31 months; I replaced the switch, invoiced for it, and left the sealed cabinet exactly as it was. The replacement lasted 14 months. My error, not the product's.
What I do now, roughly in order of effect per dollar. Vent the box and filter the vent — two 60 mm filtered openings on opposite faces turn a sealed oven into a cabinet, and it is usually the biggest single lever. Leave a 1U gap above the switch. Aim for intake under 40 °C; treat 35 °C as the target rather than the stretch goal.
Keep PoE load at or below about 70% of nameplate in a closed box. A 48-port switch at 40% load runs cooler and lasts longer than a 24-port at 90%, and it costs less than redesigning the cabinet's thermal envelope. For small run counts, an external adapter moves the conversion loss out of the box entirely — which is why single-port injectors survive so well in hot corners despite their other limits. Where the environment is genuinely hostile — unventilated plant rooms, roof cabinets, anything above 45 °C for hours a day — the honest answer is a unit built for it rather than a commercial unit with a bigger heatsink. That is the difference between a switch rated for the temperature and one merely rated for the port count.
Questions I Get Asked About PoE Switch Longevity
Does a hotter switch always mean a worse one? No. Efficiency and load fraction matter more than surface temperature. A 90%-efficient supply at 70% load dumps less heat than an 82% one at 95%, even with the bigger case. Compare watts of loss, not degrees on the lid.
Do solid polymer capacitors solve it? They remove the drying-out mechanism, which is real progress on the port converters. But bulk energy storage on a 54 V rail still usually means electrolytic. An improvement in one place, not a cure for a cabinet at 50 °C.
Should I turn PoE ports off at night to extend life? Not for cameras. The saving is small, and you trade steady heat for thermal cycling plus a daily inrush event. For access points on a schedule it is defensible; for anything with a mechanical or IR load, leave it running.
How long should a 24-port PoE switch last? Seven to ten years in a ventilated closet near 30 °C. Three to four in a sealed wall box at 45 °C. If yours died at three, the box is the first suspect and the brand is the second.
Does a bigger power supply run hotter? Usually not. A supply at half load is more efficient and cooler than a smaller one pinned at 90%, with headroom left for a hot afternoon. Oversizing the PSU is one of the cheap wins here.
What I keep coming back to is how ordinary the failure is. A switch that dies at 34 months in a 45 °C cabinet did not fail early — it failed on time, and the arithmetic to know that was available on day one. If you are replacing one this month, measure the box before you order and put the reading in the handover file. If your ripple numbers tell a different story than mine, I would like to see them — the 100 Hz growth curve is the part of this I have the least data on.