A PoE switch datasheet has two performance numbers printed three lines apart, and they answer completely different questions. Switching capacity, in Gbps, is how much data the fabric could theoretically move. Forwarding rate, in Mpps, is how many packets per second the chip can actually look up, queue and hand on. Most buyers read the first, because it is bigger. The second is where the truth lives, and it is the one that gets quietly rounded down when a product is built to a price.
Last week I had a 24-port unit on the bench in Shenzhen with 56 Gbps on the label and 41.7 Mpps next to it. My tester pushed 64-byte frames into all 24 copper ports and the counter stopped climbing somewhere around 24 Mpps, with the loss LED blinking before I got to the third measurement. Power was perfect the whole time. All 24 ports held their class, the PoE budget read exactly what the spec sheet said, and the switch still could not forward what it claimed. That gap between the label and the silicon is the most useful thing I check before a container leaves the factory.
Why the Two Headline Numbers Never Agree
Switching capacity is arithmetic. You take the port count, multiply by the port rate, multiply by two because the marketing convention counts a full-duplex link as both directions at once, and you print the result. Nothing in that calculation has ever been near a packet. A 24-port gigabit switch with four gigabit SFP uplinks is 28 ports × 1 Gbps × 2 = 56 Gbps. Every vendor selling that port layout will print 56 Gbps, and they will all be right, and the number will tell you nothing about how the device behaves under load.
Forwarding rate is supposed to be the measured counterpart. It is quoted in millions of packets per second, and the honest way to get it is to flood the switch with minimum-size frames and count what comes out the other side. The dishonest way is to calculate it with the same pencil that produced the Gbps figure, which is what happens more often than anyone admits. When both numbers are derived from the port list rather than measured, the datasheet is a description of a connector layout, not a switch.
Here is the tell. Take the forwarding rate, multiply each gigabit port by 1.488, and see whether the total matches the port count exactly. If it does, to two decimal places, nobody measured anything. A measured figure almost never lands on the theoretical value — it lands a bit under, and the good vendors print the measured one anyway.
What 1.488 Mpps Actually Is
The constant shows up in every Chinese and Taiwanese datasheet and almost never gets explained, so here is the derivation. Ethernet has a fixed cost per frame that has nothing to do with payload size: 8 bytes of preamble and start-frame delimiter, plus a 12-byte inter-frame gap. Add the 64-byte minimum frame and every small packet occupies 84 bytes of wire time.
- Gigabit line rate: 1,000,000,000 bits/s ÷ 8 = 125,000,000 bytes/s.
- Bytes consumed per minimum frame: 64 (frame) + 8 (preamble) + 12 (gap) = 84 bytes.
- Packets per second at line rate: 125,000,000 ÷ 84 = 1,488,095 pps, or 1.488 Mpps per port.
- Scale for other rates: 100 Mbps → 0.1488 Mpps. 2.5 Gbps → 3.72 Mpps. 5 Gbps → 7.44 Mpps. 10 Gbps → 14.88 Mpps.
- Whole switch: sum the per-port figures. 24 gigabit ports = 35.71 Mpps. Add four gigabit uplinks and you are at 41.66 Mpps.
This is why small packets are the only honest test. At 1500-byte frames a gigabit port only has to handle about 81,000 packets per second, and almost any gigabit MAC manages that. At 64 bytes the same port needs 1.488 million lookups, and that is where the $12 switch and the $120 switch stop being the same product. Video is mostly large frames, so a camera network can look fine for years and then fall over the day a chatty discovery protocol is enabled.
The Install Where a "56 Gbps" Switch Dropped Everything
I want to be precise about this one because I got the diagnosis wrong twice. A warehouse in Dongguan, 38 cameras, two 24-port PoE switches, a single NVR. The complaint was that cameras kept vanishing from the recorder for a few minutes at a time, usually in the afternoon. Nobody could reproduce it on demand.
My first theory was the cameras. I swapped two of them, updated firmware on a third, and wasted a day. My second theory was the cable run to the far aisle, because it was 92 m and I had already been burned by a bad termination that month. I re-terminated both ends and the dropout came back the next afternoon, at almost the same time.
What actually happened was duller. Someone had patched a spare cord from a wall port back into a second wall port in the same cabinet, on the same switch, creating a loop. The switch had no spanning tree. Broadcast frames multiplied, the fabric saturated with minimum-size packets, and the management plane starved. The PoE side never noticed: every port kept its 48 V, every camera kept drawing its 9 W, and the link LEDs stayed green. Data died, power did not.
Could a bigger Gbps number have saved it? Not for a second, and I have made the same argument about uplinks before. The loop generated a storm that any 1 Gbps link will happily carry. What would have helped was a forwarding rate high enough to keep the control plane alive under storm load, a buffer that could absorb the burst instead of dropping management frames first, and the broadcast storm control that nobody had turned on. The label said 56 Gbps. The box behaved like a device that had never been asked the question.
Full Duplex, Doubling, and the Fiction of "Switching Capacity"
The ×2 convention deserves its own paragraph, because it is the source of half the confusion in this category. A switch does not forward 2 Gbps through a gigabit port. It receives 1 Gbps and transmits 1 Gbps, and those are the same frames making one trip. Counting both directions doubles every number on the page without moving a single extra packet. That is why a 24-port gigabit switch can be advertised at 48 Gbps and still have less real throughput than a 10 Gbps uplink.
Some vendors now print "switching capacity" without the ×2, which makes an identical box look weaker than a competitor's. Others print "backplane bandwidth" for a fixed-configuration switch that has no backplane at all — the term belongs to chassis systems with line cards, and on a 1U box it is decoration. I have stopped arguing about vocabulary. I ask one question instead: is the forwarding rate at least the sum of the ports × 1.488?
| Port layout | Capacity with ×2 (Gbps) | 64-byte wire rate (Mpps) | What a padded datasheet prints |
|---|---|---|---|
| 8 × 1G | 16 | 11.9 | 16 Gbps, 8 Mpps |
| 16 × 1G + 2 × 1G SFP | 36 | 26.8 | 36 Gbps, 14 Mpps |
| 24 × 1G + 2 × 1G SFP | 52 | 38.7 | 52 Gbps, 20 Mpps |
| 24 × 1G + 4 × 1G SFP | 56 | 41.7 | 56 Gbps, 24 Mpps |
| 48 × 1G + 4 × 10G SFP+ | 176 | 130.9 | 176 Gbps, 71 Mpps |
Read the right-hand column as a warning sign, not a specification. A device that prints 24 Mpps where the port math says 41.7 will still pass a camera deployment, because cameras do not send minimum-size frames. It will fail on a storm, a firmware push to 200 devices at once, or a multicast group nobody pruned. The failure looks like a software bug and gets escalated as one, which is how I end up reading packet captures at 11 p.m.
Where PoE Switches Get Cheated Specifically
PoE adds two wrinkles that plain switches do not have. The first is the power supply. A 24-port unit with a 370 W PoE budget and about 22 W of system load needs roughly 445 W at the AC inlet at 88% efficiency, and roughly 75 W of that difference becomes heat inside the box, in the same chassis as the ASIC — the component whose margin the vendor trimmed to hit a price. A hot chip forwards fewer packets. I have watched a unit that held 38 Mpps on the bench at 24 °C lose almost a fifth of that after two hours in a closed wall cabinet at 46 °C.
The second wrinkle is thermal. Every watt spent on PHY and magnetics is a watt the enclosure has to move. Cheap designs put 24 PoE ports, four SFP cages and a fanless heatsink in one 1U box and hope the ambient stays at 25 °C. In a wall cabinet in July it does not, and the first symptom is not a shutdown — it is packet loss on the busiest ports while every PoE counter still looks perfect.
There is also a quieter trick: quoting the budget as "24 × 30 W = 720 W" when the supply is 370 W. The port cap and the total are separate limits, and the total decides whether the last six cameras come up. That is the same habit of mind — printing a number that was multiplied rather than measured.
My Ten-Minute Datasheet Check
This is the sequence I run before I let a new model into a quote. It takes ten minutes and it has caught four mislabeled units this year, two from suppliers I otherwise trust.
- Count the ports and the rates. Write them down, including uplinks. Do not trust the "port configuration" summary line, which sometimes omits the SFP cages because they share a PHY.
- Compute the wire rate: gigabit ports × 1.488, 2.5G × 3.72, 10G × 14.88. That total is the floor a line-rate switch must meet.
- Compare with the printed Mpps. Anything below 70% of your figure is a switch that will drop under minimum-size load. Between 70% and 100% is a judgment call tied to the application.
- Check whether the capacity figure is exactly ports × rate × 2. If it is, and the Mpps is exactly ports × 1.488, the performance section was calculated, not tested.
- Look for buffer size and MAC table size. A shared 512 KB buffer is about 4 ms at gigabit ingress; 4 MB is about 33 ms. A table of 8K entries is fine for a camera VLAN and not fine for a flat /16 with 6,000 devices.
- Then confirm the thermal claim. If the datasheet says 50 °C operating but the PoE budget is only rated at 25 °C, the two statements contradict each other and the smaller one wins.
Step five is the one people skip. Buffer is the only place a burst can live while the egress port is busy, and it is the difference between a switch that absorbs a firmware rollout and one that drops the ARP replies that hold the network together. Given a choice between another 40 Gbps of paper capacity and 8 MB of real buffer on a camera network, I take the buffer every time. That is a preference, not a law, but twelve years of field calls have not moved me off it. The same discipline applies after the boxes are on the wall: six numbers recorded at handover turn a datasheet claim into a baseline you can defend.
Questions I Get Asked at the Bench
Is a higher switching capacity ever bad?
Only in the sense that it costs money and heat. Above the line-rate floor, extra capacity buys nothing on a fixed-configuration box, because the ports cannot generate more traffic than they can carry. What you want is an Mpps figure that meets the port math and a buffer sized for your burst profile. A 176 Gbps figure on a 48-port switch is normal. A 176 Gbps figure with 40 Mpps next to it is a machine that falls over the first time a loop appears.
My switch says 56 Gbps but the ports only add up to 28. Is that a lie?
No, it is the ×2 convention. Twenty-eight gigabit ports at full duplex gets counted as 56 Gbps. Both directions are counted even though a frame only crosses the fabric once. Every vendor does it, including the good ones, so treat the Gbps figure as a port inventory and the Mpps figure as the performance claim.
Does PoE power draw reduce forwarding performance?
Not electrically — the PoE rails and the switch ASIC are separate domains on any competent design. Thermally, yes. Every watt delivered to a camera is a watt the supply converted and lost part of as heat, and that heat raises the ASIC temperature. A switch rated at 41 Mpps cold will not hold 41 Mpps at 60 °C junction temperature. This is why I insist on seeing a derating curve rather than a single ambient number.
Can I measure forwarding rate without a test instrument?
Roughly, yes. Two hosts, one flooding UDP with 64-byte payloads and one counting, will show you where loss begins — but the hosts usually bottleneck before the switch does. A better poor-man's test is a loop you create on purpose with storm control off, watching whether the management interface survives. If SSH dies while PoE stays up, you found the control plane's limit. Do it on a bench, never on a live site.
Do jumbo frames change the numbers?
They change the pps requirement, not the capacity. At 9000-byte frames a gigabit port only needs about 13,600 packets per second, so a weak chip looks fine. That is why storage and video vendors love jumbo frames, and why a camera network tuned for them can hide a forwarding deficit for years.
None of this requires a lab. It requires reading two lines in the right order, doing one multiplication, and being willing to ask a supplier why the second number is 60% of the first. The answer is usually "that is the theoretical value", which is a polite way of saying the measured value was not good enough to print. I would rather know that before the container ships than after the cabinet overheats in August.