Six numbers, written down and dated, settle almost every PoE argument that starts six months after handover. Loaded voltage at the far end. Worst-case watts per port. An error-counter delta over a fixed traffic window. What the port thinks it allocated. Bond continuity. Cabinet intake temperature. Record those and you own a baseline. Skip them and you own a story about how it worked fine when you left.
Last month I stood in a riser closet with a facilities manager holding a sheet that said, in one line, "PoE tested — all ports OK." Fourteen cameras on that floor had been rebooting at dusk since early summer. He wanted to know whether the install was ever any good. I could not tell him, which is the part that still bothers me, because the information existed once. Somebody had a meter in their hand on commissioning day. They just didn't write the number down. We spent a morning re-deriving what a five-line table would have given us in ten seconds. The answer: a 61 m Cat5e run, a slightly green connector, and 43.9 V at a camera that needs 42.5 V minimum with its illuminator on. Marginal since day one, and only a fault once the ceiling got warm.
"It Works" Is Not a Measurement
I have signed off installations using exactly that sentence, so this is not a lecture from a distance. You plug everything in, sixteen tiles go green, and the project is done. Nothing failed. That feels like evidence.
It isn't. A link that works at 22 °C in March with the IR off is a link running at maybe 60% of its stress envelope. It has not been tested, it has been demonstrated. The difference surfaces later in a form that gets billed to the wrong party: a camera that reboots when its heater engages, an access point that trains down to 100 Mbps on a 2.5G port, a luminaire that flickers at 20:00 and is fine at 09:00. Three symptoms, one cause — not enough voltage at the load — and all three are invisible to the test most people run, which is "does it boot."
What follows is the sheet I use now, deliberately short. A form with forty lines gets filled with ticks and no numbers.
The Six Numbers, and What Each One Predicts
| # | Record | Where | A pass looks like | A fail predicts in month six |
|---|---|---|---|---|
| 1 | PD input voltage, worst-case load | Inline at the device | Above the PD floor (42.5 V for 802.3at/bt) with margin | Reboots at dusk or on cold mornings |
| 2 | Actual watts per port, all functions on | Switch telemetry, checked with a clamp | Below the class ceiling with ~20% headroom | Budget exhaustion at 70% port count |
| 3 | CRC delta over 10 min of full-rate traffic | Interface counters, before and after | Zero, or a rate you can explain | Video artefacts, speed renegotiation in rain |
| 4 | Allocated class vs measured draw | PoE status page or LLDP TLV | Within one class step | Phantom budget, half the switch unusable |
| 5 | Bond continuity to building earth | Two-point, long lead compensated | Under 1 Ω, short straight conductor | Surge damage that walks past the protector |
| 6 | Cabinet intake air temperature | Probe at the grille, 15:00 | Below the PSU derating knee | PoE output sagging quietly, PSU dead in year three |
Only one of those six is a power measurement in the usual sense. The rest are evidence about margin, and margin is what you hand over.
Number One: Loaded Voltage, Measured at the Wrong End
The switch will happily report 53.5 V at the port. That number is taken at the PSE, before the cable, and on its own it is close to useless. What the device sees is that voltage minus the drop across the whole channel:
VPD = VPSE − Itotal × Rchannel
A 100 m run of 24 AWG measures about 9.38 Ω per conductor. In two-pair mode the current leaves on one pair and returns on the other, both conductors of each pair in parallel, so the channel sits near 9.38 Ω. In four-pair mode the current splits across all four pairs and effective resistance halves to roughly 4.7 Ω — which is the whole physical case for 802.3bt using every pair, and why a 90 W upgrade is a cabling conversation too.
| Mode | Current at full load | Drop, 100 m 24 AWG | Drop, 100 m 23 AWG | PD volts from a 53 V PSE |
|---|---|---|---|---|
| 802.3af, 15.4 W, 2-pair | 0.35 A | 3.3 V | 2.6 V | 49.7 V |
| 802.3at, 30 W, 2-pair | 0.60 A | 5.6 V | 4.5 V | 47.4 V |
| 802.3bt Type 4, 90 W, 4-pair | 1.73 A | 8.1 V | 6.5 V | 44.9 V |
Those are cold-cable numbers with ideal connectors. Add 1–2 Ω for patch cords, terminations and the port FET. Add another 12–14% to the copper if the cable is sitting in a 50 °C void instead of a 20 °C one, since resistance climbs about 0.393% per °C. On the 90 W row that is the difference between 44.9 V and something nearer 41 V at the device: under the floor, on a link that passed every boot test you ran in spring. This is also where PoE extenders and fibre become the honest answer instead of a longer copper run.
Measure at the device, under load, with the load at its worst. On a PTZ that means self-test plus heater. On a camera, IR forced on in the web UI. On an AP, clients associated and the radio at full tilt. A reading taken with the device idle proves the cable is continuous. It says nothing about margin.
Numbers Two and Three: Watts, and the Error Delta
Nameplate is a compliance number, not a design number
I keep photos of clamp-meter readings next to spec sheets, and the pattern holds: a fixed dome rated 8 W draws 11–12 W with the illuminator on, a PTZ rated 18 W touches 26 W mid-pan, an AP rated 15.4 W pulls 19 W for a few seconds while its radios come up. None of that is a fault. It is what the hardware does.
It matters because of allocation. Most PoE switches I work with reserve statically by default: the full class ceiling per port, never revisited. Twenty-four ports at 30 W is 720 W against a 370 W budget, so the switch starts refusing ports at thirteen. I watched a project order a second switch over this. The cameras were 7 W devices reserved at 30 W. Fix took ten minutes of configuration and a table of measured draws. Record allocated and actual side by side; a gap wider than one class step is a decision, and it is far cheaper now than in month six.
Power faults arrive as data faults
CRC counters are the earliest warning you will get, because the marginal connector that costs you a volt and a half also mangles the differential signal. The method matters more than the instrument:
- Note or clear the counters on every PoE port. Screenshot them — remembered numbers are always wrong.
- Bring each PD to worst-case load: IR on, PTZ cycling, radios busy.
- Run ten minutes of real traffic at or near line rate. Ping is not traffic.
- Read again, divide by the interval. Anything past a few errors per minute on gigabit copper deserves a cable test.
- Check negotiated speed and duplex while you are in there. A 2.5G port trained at 100M has a pair fault, not a power fault.
Three bad punch-downs caught this way, all on links streaming 4K without a visible glitch. The counters were the only thing complaining.
Numbers Four, Five and Six: The Ones Nobody Writes Down
What the port believes it is doing
Handover is the last moment when the paper design and the running system are visible at once. Ask the switch what each port negotiated, what it reserved, and what the meter says. The mismatches are diagnostic. Class 4 negotiated, 30 W reserved, 7 W delivered is over-reservation — a slow capacity problem. Class 4 negotiated and 31 W delivered is a device drawing outside its class, usually after a firmware update enabled something nobody budgeted. Class 0 stuck behind a Type 2 device often means the PSE fell back after a failed detection cycle, and the cable is the suspect, not the device.
LLDP power negotiation is worth configuring where the switch supports it. A PD that asks for what it needs and gets a yes leaves no phantom reservation anywhere. Unmanaged hardware cannot do this, which is one of the real costs of buying unmanaged, and it only shows up the day you add the seventeenth camera. If the run is long enough that copper stops being sane, a media converter at each end removes the voltage-drop problem entirely rather than shaving it.
Bond continuity, because surge damage is a slow invoice
If the run leaves the building, crosses a roof or lands on a pole, this is not optional. Measure from the protector's earth terminal to the building earth bar: under 1 Ω, short and straight, no coiled green-and-yellow slack. I once measured 14 Ω on a bond that looked immaculate, because twenty metres of cable had been coiled behind a panel and that inductance made the protector decorative. Photograph outdoor connectors at handover with the date visible. It ends arguments about whether water ingress was pre-existing.
Cabinet temperature, taken on the worst afternoon you can find
Take it at the intake grille at 15:00, door closed, load running, and write the date beside it. Every PoE supply has a derating curve and most start bending around 45 °C intake. A 370 W supply that is honest at 25 °C may be a 260 W supply in a sealed wall box in July, and the switch will not announce it — it will start refusing ports according to a priority table nobody configured. Related and worth knowing before you fill a tray: TIA TSB-184-A holds cable rise to 15 °C, which assumes 45 °C ambient against a 60 °C cable rating. A 100-cable Cat5e bundle at full Type 4 runs about 22 °C above ambient. That is a design error rather than a handover finding, but the intake temperature is what makes it visible. More on bundle behaviour in the cable comparison I wrote earlier.
Where I Got It Wrong
For years my handover test was one line: unloaded voltage, read off the switch status page. Easy, and usually in agreement with reality. Then a job with twelve 90 W LED luminaires taught me otherwise. All twelve booted and dimmed. Two weeks later the far four began dropping out at sunset, when they go to full output. The switch reported 52 V on every port. At the luminaire I measured 40.1 V under load, on a 96 m run of 24 AWG through a roof void that had been 48 °C all afternoon.
My first call was the driver, and I said so out loud, and I was wrong in a way that cost the driver vendor a week of email. The cable was the problem. So was I, for signing a sheet with a number taken at the wrong end of the channel. We moved those four runs to 23 AWG and the drops disappeared. The lesson I actually keep is narrower and less flattering: an unloaded voltage reading mostly proves the port can source current into nothing.
Questions That Come Up at Every Handover
Can I just read the switch's power page and skip the meter?
Read it, but don't rely on it alone. The switch measures its own output, not the device's input. On a 90 m run that gap is 8–10 V — precisely the part you need. A cheap inline meter between patch cord and device closes it.
How long should the traffic window be?
Ten minutes at full rate surfaces a marginal connector. It will not surface a thermal problem; for that you need a warm cabinet and maximum load, which is why I schedule at least one handover visit for the afternoon. If the site has a dusk step-up in load, be there for it once.
Do the same six numbers work at 90 W?
They work but they are not enough. Add per-pair current balance, because four-pair delivery with one high-resistance pair loads the other three harder while the total still looks fine. Add a bundle temperature check. At that power level the cable is the limiting component, not the switch.
Is a 5% voltage drop acceptable?
Wrong question, and I hear it constantly. Compare against the device's minimum input, not nominal. 802.3af PDs are specified to 37 V; 802.3at and 802.3bt PDs to 42.5 V. A 6 V drop from 53 V is comfortable for the first and thin for the second. The percentage means nothing without the floor.
What if I inherit a site with no records?
Build the baseline backwards. Take the three longest runs and the three highest-power devices, measure those six, and date the sheet as retrospective. Not as good as a commissioning record, but it turns "it used to work" into a number, and a number can be argued with.
Who keeps the sheet?
Whoever answers the phone in eighteen months, which in practice is the integrator. Send a copy with the as-built drawings and put the counter screenshots in the same folder. The photograph of port 14's counters on handover day has settled more disputes for me than any report I have written.
None of this is difficult instrument work. A meter, a clamp, a laptop, a thermometer, forty minutes. What it buys is a defensible position: when a camera starts rebooting in August you compare against a number instead of a memory, and you know within minutes whether you are looking at a device, a cable or a cabinet.
If you have a handover sheet you are proud of — or one you are embarrassed by, which is more useful — I would like to see the columns. I am still arguing with myself about whether port-level temperature earns a seventh line, and whether the error window should be ten minutes or thirty. Real numbers from real sites settle these things faster than I can reason about them alone. Older write-ups in the PoE knowledge section cover the failures these measurements are meant to catch.