A PoE switch rated for 50 °C ambient will sit happily in a 52 °C wall cabinet for about two hours, then start refusing to power the heaviest port on the unit. Nothing is broken. The switch is doing exactly what its datasheet promised, in a box no datasheet ever described.
I opened a cabinet in a Futian corridor last Tuesday. Twenty-four ports, eleven cameras, one access point, a router and a media converter in a 500 × 400 × 200 mm steel box, door shut since 2023. The room was 31 °C. The air two centimetres above the chassis read 49 °C. The port carrying the 802.3bt access point had logged fourteen "insufficient power" events in nine days, every one between 13:10 and 16:40. The integrator had replaced the access point twice. The problem was sheet metal.
The Label Was Measured in a Room, Not in a Box
Every thermal number on a datasheet comes from a bench with free air around the chassis, usually 1.5 m of clearance all round. The 50 °C limit means the switch survives when the air entering its vents is 50 °C. It says nothing about what that air has to be for the switch to be the only thing heating it.
Inside a sealed cabinet, a PoE switch is not a network device with a temperature. It is a heater that also forwards frames. Everything it draws from the mains and does not send down a cable leaves as heat in that box, and the box has one way out: its surface.
Why a PoE switch is a heater with ports
A data-only 24-port switch draws maybe 20 W and turns nearly all of it into heat. A PoE switch with a 370 W budget draws up to about 420 W, sends 370 W down the cables to devices in other rooms, and turns the rest into heat in the chassis. At 88 % efficiency that is roughly 50 W at full load, 25 W at half load, and still 10 to 15 W at idle. Those numbers decide the cabinet, and almost nobody measures them before choosing a box.
The two numbers people read as one
Two separate thermal limits live in a PoE switch and the industry is careless about the difference. The first is operating ambient — above it the unit may shut down. The second is where the budget starts to derate, where firmware decides it can no longer guarantee class allocation. On my units that second number runs 5 to 10 K below the first: 370 W at 25 °C, about 260 W at 50 °C. A 30 % cut in a footnote nobody reads until a camera fails.
Some cheap units publish no derating curve at all — one wattage, one temperature range, and you are left to assume they apply simultaneously. They do not. If a datasheet gives a budget with no temperature qualifier, treat it as a 25 °C number and knock 30 % off for a cabinet in summer. Crude, but it works.
Do the Heat Balance Before You Drill Anything
You do not need a thermal chamber. A sealed enclosure reaches a steady state where heat in equals heat out, and the arithmetic is one line: ΔT = P ÷ (U × A). P is heat dissipated inside the box in watts, A is the surface that can lose heat in square metres, and U is how readily it sheds heat to still air — about 5.5 W/(m²·K) for painted steel, 3.5 for ABS, less when the box is recessed into a wall.
Worked through for the Futian cabinet. A 500 × 400 × 200 mm box has 0.76 m² of surface; flat on a wall, the back face is useless, so call it 0.55 m². Painted steel gives U × A = 5.5 × 0.55 ≈ 3.0 W/K. One number tells you everything: every 3 W of internal loss buys 1 K of rise.
| What is inside the box | Heat staying in the box | Rise over the room | Cabinet air at 33 °C room |
|---|---|---|---|
| Switch idle, no devices attached | 12 W | 4 K | 37 °C |
| Half the budget in use (≈ 185 W to devices) | 25 W | 8 K | 41 °C |
| Full 370 W budget in use | 50 W | 17 K | 50 °C |
| Full budget plus a router and a media converter | 62 W | 21 K | 54 °C |
Two things fall out. The rise is set by load, not by weather — the same 50 W that is fine at 22 °C room in March is a problem at 33 °C in July. And the door is the whole story: open, the intake read 34 °C; shut, forty minutes later, 47 °C. Sheet steel with no vents is an excellent lid.
Run the same numbers for a 400 × 300 × 150 mm ABS box, the one people buy for a small camera cluster because it is cheap and tidy. Effective area 0.33 m², U = 3.5, so U × A ≈ 1.16 W/K. Fifty watts of loss is a 43 K rise: at 33 °C room, 76 °C inside. That is not a derating problem, that is a fire waiting for a reason. Plastic is for splitters and media converters, not switches.
The half of the arithmetic everyone skips
Air stratifies, and cabinets are tall while switches are wide. In a 400 mm box the top runs 3 to 5 K warmer than the bottom, and the intake sits at one end of the chassis. Mount the switch at the top and you have handed it the hottest air in the box; mount it low with the router above and it inhales the coolest air available. Free fix. Nobody does it.
What a Cut Budget Looks Like From the Camera End
Derating is not a graceful slowdown. When firmware decides the thermal envelope no longer supports what it promised, one of three things happens, and none of them look like a heat problem from a laptop.
The first is refusal. A Class 8 device asks for 90 W of allocation. If the budget has been trimmed to 260 W and eleven other ports hold their class reservations, the switch may decline that port entirely — a detection pulse and nothing else. On site that looks like a dead port, so the integrator swaps the device. Twice, in Futian, while the fault kept following the clock.
The second is demotion. The port comes up at a lower class, the device boots, then browns out under load. This is the confusing version: link up, log clean, customer says "the camera works but drops out". Same mechanism as the class negotiation trap, different trigger.
The third is shedding. A managed switch with priorities configured drops the lowest-priority ports to protect the rest — that is what a priority design is for. Configure nothing and the firmware picks, and its opinion of which camera matters is not your opinion.
The diagnostic fingerprint is the clock, not the cable. A site that fails between early afternoon and late evening and recovers overnight has a thermal budget problem. A bad run fails when it rains, when a door slams, or at random. Heat keeps office hours.
Six Fixes, Ranked by What They Actually Change
I have tried all of these. Ranked by effect per unit of effort, not by price.
| Fix | Effect on the heat balance | Cost and effort | When it is right |
|---|---|---|---|
| Move the switch out of the box | Removes the problem instead of mitigating it | Small rack or shelf, plus cable | Whenever there is anywhere else to put it. Almost always. |
| Louvres top and bottom of the door | Raises effective U by about a third; ΔT drops 25–30 % | Hole saw and a filter kit | Dust is not a concern and the room is ventilated |
| Thermostatic filtered fan kit | Effective U roughly doubles; ΔT can halve | Fan, filter, thermostat, mains feed — and a filter you actually change | Box must stay sealed and dust is moderate |
| Split the load across two switches | Halves the loss per box — and doubles the boxes | A second switch and cabinet | Two clusters in different parts of a building anyway |
| Cut the per-port class | Removes watts, so removes conversion loss | Config change and a hard conversation | The 90 W devices do not need 90 W |
| Move the conversion outside | Takes supply loss out of the box | External midspan, one more device to fail | One high-power device, box already full |
On the "move it out" line I will be blunt, because I have argued about this on site. A DIN-rail industrial switch rated to 65 °C is not a solution to a sealed box that reaches 70 °C. It is a more expensive way to be outside the specification. Buy wide-temperature when the ambient is genuinely wide — a plant room, a sun-shielded roadside cabinet — not as a substitute for ventilation.
The Test I Run Now, Every Time
I stopped trusting arithmetic alone after getting one badly wrong. This takes two hours and one thermometer, and it has caught three cabinets the maths said were fine.
- Load the box for real. Every device connected and drawing, every PoE port powered, door shut. A test at 20 % load proves nothing about August.
- Probe the intake, not the room. 2.5 cm from the chassis intake, on the side the fan pulls from. Room temperature is irrelevant; intake is the number the datasheet is about.
- Probe the top of the box too. The gap tells you how much stratification you live with, and whether dropping the unit 200 mm is worth ten minutes.
- Wait for steady state. Forty minutes gives a trend, two hours gives a number. The first reading after power-up understates the final one by 8 to 12 K, and that understatement is the margin you were counting on.
- Compare against the curve, not the headline. If the manual lists only a maximum ambient, apply the 30 % haircut and see whether the real load still fits. If not, you found the fault before the customer did.
- Write it down and leave it in the cabinet. Date, room temperature, intake temperature, load. Cheapest diagnostic tool in the building, and the next person to open that door will not have to guess what normal was — same reason I record six numbers at handover.
One note on instruments. An infrared gun pointed at a black chassis reads the surface, not the air, and a working switch runs 15 K hotter on the case than at the intake, because that is where the heat leaves. Good for finding a hot spot, useless for deciding whether the unit is inside its rating. Two people once argued for twenty minutes over two IR readings that were both correct.
Questions I Get Asked About Hot Cabinets
Is 50 °C inside a cabinet a problem if the switch is rated for 50 °C? Not at exactly 50 °C. It is a problem at 50 °C plus the 2 or 3 K stratification adds at the intake, plus a budget that started derating at 45 °C. You are not at the limit, you are past it. Treat 45 °C cabinet air as the working ceiling for commercial units, 55 °C for industrial.
Can I just leave the door off? For a week, to prove the diagnosis, yes. Permanently, no. If removing the door fixes it you have learned that the enclosure is the problem, which is useful. Now solve it properly.
Why does it only fail in the afternoon? The box is a thermal integrator with a two-hour time constant, so it does not track the room instantly. By early afternoon the room has been warm for hours, the cabinet has caught up, the budget has derated and the heaviest port is the first casualty. By midnight everything powers up again, which is why the morning check passes.
Does a bigger PoE budget help? No. A bigger budget means a bigger supply and more conversion loss in the same box, so it usually makes the thermal problem worse while solving the electrical one. Two 12-port switches in two places beat one 24-port in a cupboard.
Is a fanless switch better in a sealed box? No, and this catches people. Fanless means heat must leave through the chassis into air that is not moving. In a sealed box that air is already hot and stays hot. Fanless removes a failure point and the only thing stirring the air. Right for open air, wrong for a closed box. And every 10 K of sustained heat roughly halves capacitor life, which is why switches die in year three.
What I Got Wrong for Two Years
For two years I was convinced this was a power supply problem. When a switch refused a port at 3 p.m. and accepted it at 8 a.m., the supply was the part behaving differently, so the supply was the suspect. I replaced supplies. I swapped brands. I wrote off two units as marginal and stopped using that model, which was unfair.
What changed my mind was a cheap thermometer and a bad afternoon. I put a probe in the intake of a cabinet that was "definitely ventilated" — it had a vent, so it counted — and watched it climb from 33 °C to 51 °C over ninety minutes while the load never moved. The supply was fine. The box was a solar oven with a switch inside.
Since then I care far less about which switch somebody bought and far more about where they put it. I test PoE hardware at a factory in Shenzhen, so my days are spent in rooms held at 25 °C with air moving — a terrible place to develop intuition about cabinets. The field keeps correcting me. Every time I decide a thermal design is adequate, somebody bolts a steel box to a west-facing wall.
Short version: do the one-line heat balance before buying the box, assume the published budget is a 25 °C number, mount the switch low rather than high, and spend two hours with a probe before handover. The watts are usually fine. The box is what fails.
If you have intake readings from a sealed cabinet — room temperature, box temperature, load at the time — I collect them. The useful range is narrow and the field data is thin. One more number from a hot corridor in August beats another datasheet curve.