If you want to know whether a PoE installation is really redundant, don't count power supplies. Count the seconds between the breaker tripping and the last camera sending a frame again. On most of the sites I visit, that number lands between 90 and 150 seconds — on switches with two PSUs, two cords and a ring to the core. The hardware is fine. The arithmetic was never done.

Last month I stood in a distribution warehouse outside Dongguan looking at exactly that. Twenty-four ports, two power supplies, two cords, RSTP ring back to the core cabinet. At 09:41 a floor polisher tripped a 16 A breaker and both cords died together, because both of them landed in the same panel, one way apart. The switch was forwarding again at 09:42:14. The last camera re-registered with the NVR at 09:43:31. One hundred and ten seconds of missing video, from a design that four people in the kickoff meeting had described as redundant.

That gap is not a product defect. It is what "redundancy" means once you follow the power down to the port.

Two cords, one breaker

The first thing I do on any site claiming dual power is trace both cords with my hand. Not on the drawing — with my hand, until each one ends at a breaker. I have found cords from the same switch going into two sockets of the same PDU strip, two sockets of the same UPS, and twice into two ways of the same panel that were bridged downstream by a single 30 mA RCD. All three had a commissioning report that said "dual feed".

Here in Shenzhen the failure mode is usually duller than a storm. A cleaner's machine, a kettle in the guard room, a compressor starting on the same phase. The breaker does what breakers do. Then the interesting part: if the site has a generator or an ATS, the transfer is not instant. A mechanical transfer switch I logged last spring took 9 seconds to move — normal for a building, fatal for a switch with 16 ms of hold-up time. The UPS covers that window, if it sits on the switch side of the ATS and not behind it.

So the checklist item is not "two feeds". It is: two feeds, two breakers, two paths to the utility, and a UPS that covers the transfer time of whatever sits between them. Anything less is one feed with two plugs.

What a second power supply actually buys you

A dual-PSU switch protects you from one thing: a supply failing internally. Diode-OR'd outputs, load sharing at roughly 50/50, hold-up of 10 to 20 ms in the bulk caps. If one supply drops, the DC rail sags for a few milliseconds and the other picks up the whole load. If the switch is well built, nothing reboots. That is a real benefit and I spec it on every cabinet that matters.

It stops being a benefit the moment the survivor cannot carry the load. This is where the datasheet and the purchase order drift apart, because "dual PSU" gets read as "twice the power" when it usually means "two of whatever we were going to fit anyway".

Load case DC load at the switch Capacity left after one fails What happens
2 × 150 W, 25 °C 409 W 150 W Survivor current-limits, rail collapses, full reboot
2 × 450 W, 25 °C 409 W 450 W Holds. This is the intended case
2 × 450 W, 45 °C cabinet 409 W 360 W at 80 % derating Short by 49 W. Holds for a while, then shuts down on over-temperature
2 × 550 W, 45 °C cabinet 409 W 440 W Holds, with 31 W of margin

The 80 % figure in the third row is not a rule of thumb I invented; it is what most supplies do above 40 to 45 °C intake. If the cabinet is in a corridor in August, the number on the label is optimistic by the time the second supply becomes the only supply. I learned that one the expensive way in 2019: a 2 × 300 W unit in a sealed wall box that "passed" load bank testing on a bench at 22 °C and browned out at 46 °C the following July.

The controller reboots even when the power doesn't

Here is the part that surprises people. On a managed PoE switch, port power is not a wire — it is a state machine. The PSE measures for a 25 kΩ signature between 2.7 and 10.1 V, classifies the PD, then ramps 44 to 57 V onto the pair. That sequence runs per port, in software-scheduled order, after the switch has booted far enough to have a PoE controller.

Which means every reboot is a full power cycle for every powered device. Not a glitch — a cycle. Ports come up in sequence over 5 to 20 seconds depending on firmware, deliberately staggered so 24 PDs don't hit the rail with inrush at once. Then each camera boots. Then each camera re-registers with the NVR.

In the warehouse that arithmetic was: 52 seconds from breaker reset to forwarding, 14 seconds to the last port being powered, then 45 seconds for the slowest PTZ to boot and re-subscribe. 111 seconds, and only one of those three numbers had anything to do with the power supply.

A few enterprise lines ship features that address this — power within seconds of AC return, before the OS is up, and power held across a software reload. They work, and on evidence-grade sites they are worth asking for by name. They do not help if the PSE itself failed.

RSTP saves the path. The camera is still rebooting.

Rapid spanning tree does one job well: when a link or a switch disappears from the topology, it finds another path. On 802.1w with a simple topology, that is typically a few hundred milliseconds. Legacy STP, if someone left it on, is 30 to 50 seconds of listening and learning. Vendor ring protocols do 20 to 50 ms, but only on the same vendor's hardware at both ends, configured as a ring — not on a general RSTP mesh.

None of those numbers describe what happens to a camera. RSTP protects the path; it says nothing about the power at the far end. In the warehouse the switch itself lost power, so the ring was irrelevant — there was no path because there was no switch. And even when RSTP does its job perfectly, an RTP video stream that loses two seconds gets torn down by the recorder, and re-establishing RTSP plus waiting for the next keyframe adds a few more seconds of unusable footage. With a 2-second GOP, you can add 2 more.

I got this wrong for years. I used to size redundancy by drawing the ring and calling it done, because a ring is visible on a drawing and a boot sequence is not.

Layer Covers Does not cover Recovery
Dual PSU One supply failing inside the box Upstream loss, undersized survivor 0 ms if sharing, reboot if not
Second feed One breaker, one PDU, one cable Both cords on the same way 0 ms
UPS Outages up to its runtime ATS transfer beyond hold-up, exhausted battery 0 ms until it ends
RSTP ring One link, one switch in the path Loss of the local switch, PoE controller faults 200 ms to 2 s, ports stay up
Held port power A control-plane reload PSE hardware failure, full power loss 0 s for power, data as usual
Ports split across two switches Loss of one switch or one PSE bank Shared upstream, shared NVR Half the cameras keep recording

A worked example: 22 cameras, 480 W, two supplies

This is the arithmetic I now do on paper before quoting anything, in the order that matters. It takes ten minutes and it has killed more bad designs than any datasheet argument.

  1. Add the PD load at the port, not at the plug. 16 domes at 11 W = 176 W. 6 PTZ at 25.5 W = 153 W. Total 329 W at the devices.
  2. Add the cable loss back in. 12.5 Ω is the standard channel budget. Six PTZ at 0.6 A: 0.6² × 12.5 = 4.5 W each, 27 W. Domes at ~0.25 A: 0.25² × 12.5 = 0.78 W each, 12.5 W. PSE output ≈ 368 W.
  3. Add the switch itself. ASIC, fans and PHYs on a 24-port managed unit: 35 to 40 W. Call it 38 W. DC load ≈ 406 W. A 370 W PoE budget switch is already out of its depth here, which is the first thing the brochure hides.
  4. Pick supplies for N+1, not for the total. Each supply must carry 406 W alone. 2 × 450 W is the smallest honest answer.
  5. Derate for the cabinet. At 45 °C intake, 80 %: 450 × 0.8 = 360 W. Now the survivor is 46 W short. Either go to 2 × 550 W (440 W usable) or drop two PTZ onto a second switch.
  6. Check the AC side. 406 W at 88 % efficiency = 461 W, about 2.0 A at 230 V. Two cords on one 16 A way will never trip on current, which is exactly why nobody notices they are one feed.
  7. Time the transfer. Ask the electrician what the ATS does and put a stopwatch on it during commissioning. Nine seconds is normal. That is 560 hold-up periods. The UPS has to be downstream of the transfer.
  8. Count the dark seconds. Boot 52 s + port stagger 14 s + slowest PD 45 s = 111 s. If the project cannot accept 111 s, no power supply on the market fixes it. Only splitting the ports can.

Notice what step 4 does to the phrase "dual power supply". Two supplies are not redundancy. Two supplies each of which can carry the whole load, at the temperature the cabinet actually reaches, are redundancy. The rest is a spare part that arrives too late.

Split the ports, not just the feed

If the requirement is continuous recording rather than a warm feeling at the kickoff meeting, the only design that delivers it is two switches. Split the cameras odd and even — not by floor, because a fire or a flood takes a floor. Each switch gets its own supply, its own feed, its own UPS if the budget allows. The recorder gets two NICs, or you run two recorders and accept the extra license cost.

It costs more than a second PSU. It is also the only arrangement in which a dead PSE bank leaves half your cameras recording instead of none. When I put the two options side by side for a client last year — one 24-port switch with 2 × 550 W, versus two 16-port switches with 1 + 1 each — the difference was about the price of two of the PTZ cameras on the same job. They took the two switches.

Two things if you go that way. First, put the split on the recorder side as well, otherwise the NVR is the single point you just paid to remove. Second, set PoE port priority on both switches before the first power cut, because the surviving switch will be running near its budget and the order in which it sheds ports is a decision someone should make on purpose. I wrote about that shedding order separately.

The last thing: redundancy you cannot see is not redundancy. A second supply that failed in March and nobody noticed is a single supply in November. If the managed switch can send a trap or close a dry contact on PSU failure, wire it to something that wakes a human up. If it cannot, check the supplies at every service visit.

Questions I get asked after every outage

Does a dual-PSU switch keep the cameras powered if one supply dies?

Yes, if the surviving supply can carry the full DC load at the cabinet's real temperature. If the two supplies were sized as a pair for the total, the survivor goes into current limit on the first cold morning with heaters on, the rail collapses, and everything reboots anyway.

How long does RSTP actually take?

On a simple 802.1w topology, a few hundred milliseconds, occasionally a second or two if the topology is deep. Vendor ring protocols hit 20 to 50 ms, but only in their own ring. Legacy STP is 30 to 50 seconds and should not exist on a video network in 2026. In all of these cases the camera keeps power — the ring is about the path, not the port.

Will cameras stay powered through a firmware upgrade?

Usually not. A normal reload drops the PoE controller and re-runs detection on every port. Some platforms offer a mode that holds power across a software restart, and it is worth testing rather than assuming: I have seen it protect 4 ports out of 16, because only ports already classified when the reload began were covered.

Can I just put both cords on the same UPS?

You can, and it beats two wall sockets on the same way. But be clear about what you built: one UPS, now the single point of failure, with a battery that is flat in 20 minutes. Two UPSs on two circuits is the arrangement that survives an outage.

Is a redundant ring enough for a 40-camera site?

For cable damage, mostly yes. For the failures I actually get called about — a tripped breaker, a dead PSU, a PoE controller that stops classifying after a bad upgrade — the ring contributes nothing, because the camera's own switch is the thing that died. Ring plus split ports plus monitored supplies is the combination that stops the phone call.

I still specify dual supplies on every cabinet that matters, and I still draw the ring. What changed is the sentence I put under it: redundancy here means the switch keeps running, not that the cameras keep recording. Those are two different numbers, and only one of them shows up on a quotation.

If you have timed your own dark seconds — boot, stagger, PD registration — I would like to see the breakdown. Mine all come from 24-port managed units in 40 to 45 °C cabinets, and a fanless 8-port at the edge probably looks nothing like it.