On a fibre link, transmit and receive power are the first numbers to look at: they tell you whether the light is leaving one end and arriving at the other with enough margin to decode. This guide covers the units, the link budget, the digital diagnostics the transceiver reports, and the fault patterns those readings produce.
Overview: the two numbers that decide whether a link works
A fibre link carries light. The transmitter turns bits into pulses of light, the fibre carries them, and the receiver turns them back into bits. Two numbers describe that journey:
Tx power (transmit power, or launch power) is how much optical power the transmitter puts into the fibre.
Rx power (receive power) is how much of it arrives at the far end.
Everything between the two ends subtracts from the first number to give the second. Fibre absorbs light. Every connector and splice loses a little. Dirt on a ferrule loses a lot. If enough is lost, the receiver cannot tell a one from a zero reliably, and the link either drops or runs with errors.
The reason these two numbers matter more than any others is that they are measurable, they are reported by the transceiver itself, and they fail in a readable way. A link that is down because the fibre is broken looks different from a link that is down because the far-end laser is dead, and both look different from a link that is up but dropping frames because a connector is dirty. You can tell which is which from the power readings alone, usually in under a minute, without a technician walking to the other end of the building.
What this guide covers
This guide is written for someone who manages switches, routers and servers and who reads these numbers off a command line rather than off a power meter. It covers:
The units, because optical power is measured in a logarithmic scale that trips people up the first few times.
What transmit and receive power actually are, and the range each transceiver type is specified for.
The link budget: the arithmetic that says whether a given run of fibre will work with a given pair of optics.
Where the readings come from inside the transceiver, how accurate they are, and what the alarm thresholds mean.
What each common fault looks like in the numbers.
Throughout, the numbers quoted for specific transceivers come from published vendor data sheets, and the cabling loss figures come from the cabling standards. Both are listed in section 15. Your own optics may differ, and their data sheet is the authority for them.
The units: dBm, dB and milliwatts
Optical power is a quantity of light, measured in watts. On a network link the quantities are small, so the natural unit is the milliwatt (mW) or the microwatt (µW). But power levels are almost always quoted in dBm, and losses in dB, because the logarithmic scale turns multiplication into addition and makes a link budget something you can do in your head.
dBm is an absolute level
dBm means decibels relative to one milliwatt. The definition is:
power in dBm = 10 x log10( power in mW )So 1 mW is 0 dBm by definition. Anything weaker than 1 mW is negative. Since transceiver powers are almost always well under a milliwatt, nearly every number you read will be negative, and a more negative number means less light. This is the single most common point of confusion: -3 dBm is strong, -30 dBm is very weak, and -30 is the worse reading even though it is the bigger number in everyday arithmetic.
dB is a ratio
dB on its own is not a power level. It is a ratio between two powers, which is how losses and gains are expressed:
loss in dB = 10 x log10( power in / power out )A connector with 0.5 dB of insertion loss removes 0.5 dB whatever the input level. Two kilometres of fibre at 0.4 dB/km removes 0.8 dB. The losses simply add up.
The arithmetic that matters
There are only four rules, and three of them are about what you may subtract from what.
| Operation | Result | Meaning |
|---|---|---|
| dBm − dB | dBm | A level minus a loss gives the level that is left |
| dBm − dBm | dB | Two levels compared give the difference between them |
| dBm + dBm | meaningless | You cannot add two absolute levels this way |
| dB + dB | dB | Losses along a path add up |
The fourth rule is the set of ratios worth memorising, because they let you sanity-check a reading without a calculator:
3 dB is a factor of two. Lose 3 dB and half the light is gone. Gain 3 dB and it has doubled.
10 dB is a factor of ten. -10 dBm is a tenth of 1 mW, so 100 µW.
They combine. -13 dBm is 10 dB then 3 dB down from 1 mW, so a tenth then a half: 50 µW.
A conversion table to keep nearby
| dBm | Power | Where you see it |
|---|---|---|
| +5 dBm | 3.16 mW | Long-haul transmitters |
| 0 dBm | 1 mW | The reference point |
| -3 dBm | 0.50 mW | Typical healthy single-mode Tx |
| -7 dBm | 0.20 mW | Typical short-reach Tx |
| -10 dBm | 100 µW | Comfortable Rx on a short link |
| -14 dBm | 40 µW | Near 10GBASE-LR sensitivity |
| -20 dBm | 10 µW | Weak; many receivers are already failing |
| -30 dBm | 1 µW | Effectively no usable signal on an Ethernet link |
| -40 dBm | 0.1 µW | The floor of the diagnostic reporting range |
That last row is worth remembering for a different reason. The digital diagnostics in the transceiver report power in steps of 0.1 µW, which is -40 dBm, so -40 dBm is the lowest non-zero value the module can report at all. A reading at or below about -40 dBm means the module is seeing nothing, not that it is seeing a very faint signal.
Milliwatts are still useful for one thing
Switch output often prints both units side by side, for example 0.3186 mW / -4.97 dBm. The linear value is the easier one when you want to compare two readings as a proportion rather than as a difference, and it is a quick way to check that a logarithm has not been mangled somewhere in a monitoring pipeline. Otherwise, work in dBm.
Transmit power: what the laser puts into the fibre
Tx power is the average optical power the transmitter couples into the fibre. On the data sheet it appears as a range, not a single figure, and the range has a floor and a ceiling for different reasons.
The minimum exists so that the link budget works. If every installed module launched at the bottom of its range and every fibre run hit its maximum allowed loss, the receiver at the far end must still get enough light. The minimum launch power is what makes that guarantee possible.
The maximum exists for two reasons: eye safety, and not overloading the receiver at the far end on a short link. Cisco's 10GBASE data sheet states this directly for 10GBASE-SR, where the launch power shall be the lesser of the Class 1 safety limit or the maximum receive power.
How the module holds the power steady
Inside the transceiver, a small photodiode behind the laser samples a fraction of the output. An automatic power control (APC) loop compares that sample against a target and adjusts the laser bias current to keep it there. As Lightwave describes the mechanism, the loop monitors the back-facet photodiode current and adjusts the bias current to keep that current constant, which keeps average power constant.
This matters for troubleshooting because of what the loop has to work against. A semiconductor laser has a threshold current below which it barely emits, and that threshold rises with temperature: the same article notes the threshold current of an edge-emitting laser can rise by more than 20 mA going from -40 to +85 degrees C. The threshold also drifts upward as the laser ages. In both cases the loop compensates by pushing more bias current through the laser to hold the same output power.
The practical consequence is a diagnostic pair:
| Tx power | Bias current | What it suggests |
|---|---|---|
| Steady, in range | Steady | Normal |
| Steady, in range | Climbing over months | The loop is compensating; worth watching, and worth checking module temperature first |
| Falling below the low threshold | At or near its high alarm | The loop has run out of range; the transmitter is failing |
| Zero or near zero | Zero | Laser is off: either disabled, or the module has shut it down |
Bias current on its own is not an alarm. Read it alongside Tx power and module temperature, because a hot module legitimately draws more bias than a cool one.
Average power and optical modulation amplitude
Data sheets usually quote average optical power, and say so. Cisco's 10GBASE and 100GBASE data sheets both carry a footnote that transmitter and receiver power is average optical power unless otherwise specified.
Some short-reach optics are additionally specified by optical modulation amplitude (OMA), which is the difference in power between a transmitted one and a transmitted zero rather than the mean of the two. OMA is the more honest measure of how easy the signal is to decode, because a signal can have healthy average power and still be hard to read if the one and zero levels are too close together. Cisco's 10GBASE-LRM entry is specified both ways, with a note that both the average and OMA specifications must be met simultaneously.
For day-to-day work you will be reading average power, because that is what the digital diagnostics report. It is worth knowing OMA exists, because it explains a link that has textbook power readings and still runs with errors: the average is fine and the separation between the levels is not.
What Tx power does not tell you
Tx power is measured inside the module, before the light reaches the connector. It is therefore silent about everything outside the module:
It does not change if the fibre is dirty, bent, broken or unplugged.
It does not change if the far end is down.
It does not confirm that the light actually got into the fibre; a badly seated or cracked connector right at the port will not move the Tx reading at all.
A healthy Tx reading at both ends and no Rx power at either end means the fibre path between them is the problem. That pairing of readings is the most useful single fact in this guide.
Receive power: sensitivity, overload and the window between them
Rx power is the average optical power arriving at the receiver. Unlike Tx power, it is a measurement of the whole path: the far-end laser, every metre of fibre, every connector, every patch panel, every splice. It is the single reading that tells you most about a link.
A receiver has two limits, and the usable band between them is narrower than people expect.
Sensitivity: the floor
Receiver sensitivity is the lowest average power at which the receiver still recovers the data at the specified bit error ratio. Below it, the photodiode and its amplifier produce noise comparable to the signal, and bits start to flip.
Two things about sensitivity catch people out:
It is a specified minimum, not a cliff. A receiver is guaranteed to work at its sensitivity figure. Many will work a little below it, and none should be relied on to. Treat the sensitivity figure as the bottom of the usable range and keep margin above it.
It is quoted at a bit error ratio. A receiver that is at its sensitivity limit is not error-free; it is meeting the error ratio the standard allows. Running a link at the limit means running it at the worst error performance the standard permits, with no margin for a connector that gets dirty next year.
Overload: the ceiling
The other limit is the maximum input power, variously called overload or saturation. Above it the receiver's front-end amplifier saturates, the pulse shapes distort, and the error ratio climbs again. As the Fiber Optic Association puts it, excess power saturates the receiver amplifier and too little power lets noise interfere with the signal, and both extremes raise the bit error rate.
Overload is the limit people forget, because it is counter-intuitive: the link has too much light and therefore does not work. It shows up in exactly one situation, which is a long-reach transceiver on a short fibre. An optic built to push light 40 or 80 km, plugged into a patch lead a few metres long, delivers far more power than its receiver was designed to take.
The window, and how wide it actually is
The difference between overload and sensitivity is the receiver's dynamic range. It is often around 15 dB, but it varies a lot by optic type, and the short-reach ones are the tightest.
| Transceiver | Rx max (overload) | Rx min (sensitivity) | Window |
|---|---|---|---|
| 1000BASE-SX | 0 dBm | -17 dBm | 17 dB |
| 1000BASE-LX/LH | -3 dBm | -20 dBm | 17 dB |
| 1000BASE-ZX | -3 dBm | -23 dBm | 20 dB |
| 10GBASE-SR | -1.0 dBm | -9.9 dBm | 8.9 dB |
| 10GBASE-LR | +0.5 dBm | -14.4 dBm | 14.9 dB |
| 10GBASE-ER | -1 dBm | -15.8 dBm | 14.8 dB |
| 10GBASE-ZR | -7 dBm | -24 dBm | 17 dB |
Values are from the Cisco data sheets listed in section 15, for Cisco-branded modules. Other vendors' equivalents are similar but not identical, so check the data sheet for the part you actually have.
The 10GBASE-SR row is the one worth noting. Its receiver has under 9 dB of range, which is less than many people assume they have to play with, and it is the most widely deployed 10G optic there is.
The bars sit at different heights as well as different widths: a 10GBASE-ZR receiver works down at -24 dBm but overloads at -7 dBm, which is why it needs attenuation on anything short.
Where in the window you want to sit
Not at either end. The FOA's advice for design is to target a receiver power near the middle of the operating range, not near sensitivity or overload, and to verify it with a power meter after installation.
The practical version for someone reading a switch CLI:
Comfortably inside both limits is healthy. There is room for the plant to degrade.
Within about 3 dB of sensitivity is thin. It will probably pass traffic today. A dirty connector or a re-patch will take it under.
Within about 3 dB of overload is equally thin, and the fix is an attenuator rather than cleaning (section 12).
Outside either limit means the link will be unreliable even if it comes up, and the interface counters will usually show it.
The link budget: where the light goes
The link budget is the arithmetic that connects the two numbers. It is one subtraction repeated:
Rx power = Tx power - fibre loss - connector loss - splice lossmargin = Rx power - receiver sensitivitywhere the sensitivity is the receiver's specified minimum. Positive margin means the link should work. Zero or negative margin means it should not, and if it appears to, it is running on tolerances that will not last.
A budget is calculated with worst-case values: the minimum launch power the transmitter is specified for, and the maximum loss each component is allowed. That is deliberate. A budget built on typical values tells you the link will probably work with the modules you have today, which is not the question being asked.
What each component costs
The cabling standard sets the ceilings. ANSI/TIA-568-C.3 gives the maximum cable attenuation and the maximum loss for connectors and splices:
| Component | Maximum loss | Source |
|---|---|---|
| Multimode fibre at 850 nm (OM1, OM2, OM3) | 3.5 dB/km | TIA-568-C.3 Table 1 |
| Multimode fibre at 1300 nm (OM1, OM2, OM3) | 1.5 dB/km | TIA-568-C.3 Table 1 |
| Single-mode inside plant at 1310 and 1550 nm | 1.0 dB/km | TIA-568-C.3 Table 1 |
| Single-mode outside plant at 1310 and 1550 nm | 0.5 dB/km | TIA-568-C.3 Table 1 |
| Mated connector pair | 0.75 dB | TIA-568-C.3 Annex A |
| Splice, fusion or mechanical | 0.3 dB | TIA-568-C.3 clause 5.3 |
Two things follow from that table that people often get wrong.
Connectors cost more than fibre on short links. At the standard's maximum, one mated connector pair costs 0.75 dB, which is the same as 1.5 km of outside-plant single-mode fibre. A structured cabling run with a patch panel at each end has four mated pairs in it before you count anything else: equipment cord to panel, panel to backbone, backbone to panel, panel to equipment cord. That is 3 dB of budget spent on connectors on a link that might be 200 m long.
Inside-plant single-mode is specified at twice the loss of outside plant. 1.0 dB/km against 0.5 dB/km. The fibre in a riser is allowed to be worse than the fibre in the ground, because the standard accounts for tight-buffered indoor cable construction.
Why multimode distance is not a loss calculation
If you budget a multimode link purely on loss, you will get the wrong answer. TIA-568-C.3 gives the same 3.5 dB/km at 850 nm for OM1, OM2 and OM3, yet the distances those fibres support are wildly different. Cisco's 10GBASE-SR entry runs 26 m on FDDI-grade multimode, up to 300 m on OM3 and up to 400 m on OM4.
The difference is bandwidth, not attenuation. The same TIA table sets a minimum modal bandwidth-length product for each grade at 850 nm: 200 MHz·km overfilled for OM1, 500 MHz·km for OM2, and 1500 MHz·km overfilled with 2000 MHz·km effective for OM3. Beyond that distance the pulses spread into each other and the link fails regardless of how much power arrives.
So for multimode, the rule is: the distance limit comes from the transceiver data sheet for your fibre grade, and the loss budget is a second check on top of it. For single-mode at these rates, loss is usually the binding constraint and the budget is the whole answer.
A worked budget, drawn
Take a 2 km campus link on outside-plant single-mode with 10GBASE-LR optics, a patch panel at each end, and one splice at a building entrance. Using worst-case values throughout:
The four connectors cost three times what two kilometres of fibre cost. That is the normal shape of a short single-mode budget, and it is why adding one more patch panel to a marginal link is enough to break it.
Where the readings come from
The Tx and Rx power you see on a switch is not measured by the switch. It is measured by the transceiver, which contains its own photodiodes, converters and a small EEPROM, and is read out by the host over a two-wire serial interface. The switch is just reporting what the module told it.
This matters because it sets the limits of what you can trust.
The memory map
For SFP, SFP+ and SFP28 modules, the relevant standard is SFF-8472, Specification for Management Interface for SFP+. It defines two blocks of memory at two addresses on the serial bus.
A0h holds identity: vendor name, part number, serial number, transceiver type, wavelength and supported reach. This is where a switch gets the text it prints when you ask what is in a port.
A2h holds diagnostics: the live readings, the thresholds they are compared against, and the flags that say which have been crossed.
The live readings sit in a contiguous block at A2h bytes 96 to 105, each a 16-bit value with the most significant byte first:
| Bytes | Quantity | Unit per least significant bit |
|---|---|---|
| 96-97 | Module temperature | 1/256 degrees C, signed |
| 98-99 | Supply voltage | 100 µV |
| 100-101 | Tx bias current | 2 µA |
| 102-103 | Tx output power | 0.1 µW |
| 104-105 | Rx input power | 0.1 µW |
The alarm and warning thresholds live at A2h bytes 0 to 55, four two-byte values per quantity in the order high alarm, low alarm, high warning, low warning. Rx power thresholds are at bytes 32 to 39, Tx power at 24 to 31.
Everything a monitoring system shows you about an optic comes out of these two blocks.
Internally and externally calibrated modules
Byte 92 of A0h carries the diagnostic monitoring type. Two bits of it decide how the A2h values must be read:
Internally calibrated (bit 5): the values at A2h are already in the absolute units above. Read and use them.
Externally calibrated (bit 4): the values are raw analogue-to-digital counts, and the host has to convert them using calibration constants stored at A2h bytes 56 to 91, including a polynomial for Rx power.
A third bit in the same byte, bit 3, says whether the Rx power figure is average power or optical modulation amplitude.
You rarely need to care, because the switch software handles it. You need to care when something downstream does not, which is the usual explanation for a monitoring system that reports a nonsensical power for one particular module type while the switch CLI reports it correctly.
Accuracy: the number that changes how you use all of this
SFF-8472 requires the Tx power and Rx power readings to be accurate to better than plus or minus 3 dB at the vendor's wavelength, across the specified temperature and voltage range. SFF-8636, the equivalent for QSFP and QSFP28, requires the same figure.
Three decibels is a factor of two. A module reporting -12.0 dBm is in specification if the true power is anywhere from -15 dBm to -9 dBm.
That has consequences worth taking seriously:
Do not certify a link against its margin using DOM alone. If the budget says you have 2 dB of margin and the module reports a reading consistent with that, you have not proved anything. A power meter is the instrument for acceptance testing; DOM is the instrument for operations.
Comparisons within one module over time are far better than the absolute figure. The error is dominated by a fixed calibration offset, so a module whose Rx power drops 4 dB over six months has genuinely lost 4 dB, even if neither the before nor the after reading is accurate in absolute terms. Trend the readings and alert on change.
Do not compare two modules' readings to each other and conclude much. Two ends of the same fibre reporting Rx powers 2 dB apart may be reading the same light.
QSFP and newer form factors
SFP-family modules use SFF-8472. Four-lane modules such as QSFP+ and QSFP28 use SFF-8636, which the specification describes as a management interface for 4-lane pluggable transceiver modules, direct attach modules and shielded cable assemblies. The concepts transfer exactly: the same 0.1 µW unit, the same plus or minus 3 dB accuracy, the same thresholds and flags. What changes is the addressing, and the fact that there are four of everything. In SFF-8636 the per-lane Rx powers sit at page 00h bytes 34 to 41 and the per-lane Tx powers at bytes 50 to 57.
The operational difference is that a four-lane module can be half broken. One lane reading -9 dBm while the other three read -3 dBm is a real and common failure, and a tool that averages the four lanes into one number will hide it. Always look at the per-lane figures.
Alarm and warning thresholds
Every monitored quantity in a transceiver carries four thresholds, stored in the module alongside the live readings. In SFF-8472 they appear in a fixed order for each quantity: high alarm, low alarm, high warning, low warning. Read as a ladder from the top down, they are high alarm, high warning, low warning, low alarm, and they cut the range into five zones.
What the levels mean
Inside the warning thresholds is the normal operating band. Nothing to do.
Between a warning and an alarm threshold means the reading has left the band the vendor considers normal but has not yet reached the point where the module is outside its specification. This is where you want to catch a degrading link: it is still passing traffic, and you have time to schedule the work.
Beyond an alarm threshold means the reading is outside the specified operating range. The link may still be up. It should not be trusted.
The flags that record these crossings live at A2h bytes 112 to 117, and the standard describes them plainly: a flag is set when the monitored value exceeds the corresponding alarm or warning level.
Who chooses the numbers
This is the part worth internalising. SFF-8472 defines where the thresholds live and how they are formatted. It does not define what they should be. The values are written into the module by whoever manufactured it.
That means:
Two modules of the same nominal type from different vendors can carry different thresholds, and both are compliant.
A threshold is not the same thing as the receiver sensitivity on the data sheet. Sometimes they line up. Often the low Rx alarm sits some way below sensitivity, so the module will not flag a link that is already too weak to work reliably.
A monitoring system that alerts purely on the module's own flags inherits whatever the vendor chose.
The robust approach is to alert on both: the module's flags, and your own thresholds derived from the data sheet. A reasonable pair of your own rules is to warn when Rx power comes within 3 dB of the data sheet sensitivity, and to warn on any change of more than 2 dB from the link's own established baseline.
A real set of thresholds
These are the Rx power thresholds reported by one module, as shown in NVIDIA's Cumulus Linux documentation for ethtool -m:
Note how much room sits between the low warning at -16.99 dBm and the low alarm at -20 dBm. On many optics the link is already unusable well before the low alarm fires, which is the argument for setting your own thresholds from the data sheet rather than relying on the module's.
Reading the values in practice
Every platform exposes the same underlying bytes. Only the presentation differs.
On a switch or router
On Cisco IOS XE the command is show interfaces transceiver detail. It prints two tables, one for transmit power and one for receive power, and each row carries the current reading followed by four threshold columns headed High Alarm, High Warn, Low Warn, Low Alarm. Cisco's troubleshooting guide for transceivers in IOS XE shows an example where the receive table carries thresholds of 2.0, -1.0, -9.9 and -13.9 dBm.
Those numbers are worth pausing on. The low warning threshold of -9.9 dBm is exactly the receive sensitivity Cisco publishes for a 10GBASE-SR module, and the low alarm sits 4 dB below it. A link on that optic reading -11 dBm has a warning but no alarm, and is already below the power the receiver is specified to work at. The warning is the line that matters.
On Cisco platforms that use the hardware module form of the command, the same data is available through show hw-module subslot <slot/subslot> transceiver <port> status, and the stored thresholds through the idprom detail variant.
Other vendors use their own wording for the same thing, commonly built around the phrase diagnostics or DOM. Whatever the command, you are looking for four things per direction: the current reading, and the four thresholds around it.
On Linux
ethtool -m enp1s0f0This reads the module EEPROM directly and decodes it. NVIDIA's Cumulus Linux documentation shows the two fields you want, printed in both linear and logarithmic units:
Laser output power : 0.3186 mW / -4.97 dBm
Receiver signal average optical power : 0.3195 mW / -4.96 dBmThe field names are not the ones you might guess:
Field in ethtool -m |
What it is |
|---|---|
Laser output power |
Tx power |
Receiver signal average optical power |
Rx power |
Laser bias current |
Tx bias current |
Module temperature |
Module temperature |
Laser rx power low warning threshold |
The low warning threshold for Rx power |
The same output carries an Optical diagnostics support field. If it says no, the module does not implement digital diagnostics at all and no power figures will appear, which is a property of the module rather than a fault.
Two practical notes. The command needs root, and it needs the network driver to implement the module EEPROM interface, so an older kernel or an unusual NIC may return nothing useful even with a module that supports diagnostics. And reading the EEPROM is a transaction on the module's serial bus; it is safe and routine, but polling every interface every few seconds is not a good idea on a busy box.
Deriving the path loss from the two ends
If you can read both ends, you can measure the fibre without a power meter:
path loss = far end Tx power - near end Rx powerDo the same in the other direction. The two results should be close, because the fibre is the same in both directions; a large asymmetry points at one specific connector or one specific patch lead rather than at the run as a whole.
Be honest about the precision. Each reading carries up to 3 dB of error, so the difference between them carries up to 6 dB. This calculation is good for answering is the path loss roughly what the budget predicted, or wildly different? It is not good for answering do we have 1.5 dB of margin? For that, use a light source and a power meter.
Published power ranges by transceiver type
The table below collects the figures from the Cisco data sheets listed in section 15. They are for Cisco-branded modules; another vendor's equivalent part will be close but not identical, and its own data sheet governs.
| Transceiver | Wavelength (nm) | Fibre and reach | Tx min / max (dBm) | Rx min / max (dBm) |
|---|---|---|---|---|
| 1000BASE-SX | 770-860 | MMF, up to 1000 m on OM3 | -9.5 / -3 | -17 / 0 |
| 1000BASE-LX/LH | 1270-1355 | MMF 550 m, SMF 10 km | -9.5 / -3 | -20 / -3 |
| 1000BASE-EX | 1290-1335 | SMF 40 km | -1 / +3 | -22 / +1 |
| 1000BASE-ZX | 1500-1580 | SMF about 70 km | 0 / +5 | -23 / -3 |
| 1000BASE-BX10-D/U | 1480-1500 and 1260-1360 | SMF 10 km, one fibre | -9 / -3 | -19.5 / -3 |
| 10GBASE-SR | 840-860 | MMF, 300 m OM3, 400 m OM4 | -7.3 / -1.2 | -9.9 / -1.0 |
| 10GBASE-LRM | 1260-1355 | MMF 220 m, SMF 300 m | -6.5 / +0.5 | -8.4 average / +0.5 |
| 10GBASE-LR | 1260-1355 | SMF 10 km | -8.2 / +0.5 | -14.4 / +0.5 |
| 10GBASE-ER | 1530-1565 | SMF up to 40 km | -4.7 / +4.0 | -15.8 / -1 |
| 10GBASE-ZR | 1530-1565 | SMF about 80 km | 0 / +4.0 | -24 / -7 |
| 25GBASE-SR | 840-860 | MMF, 70 m OM3, 100 m OM4 | -8.4 / +2.4 | -10.3 / +2.4 |
| 25GBASE-LR | 1295-1325 | SMF 10 km | -7.0 / +2.0 | -13.3 / +2.0 |
| 100GBASE-SR4, per lane | 840-860 | MMF, 70 m OM3, 100 m OM4 | -8.4 / +2.4 | -10.3 / +2.4 |
| 100G CWDM4, per lane | 1271, 1291, 1311, 1331 | SMF 2 km | -6.5 / +2.5 | -11.5 / +2.5 |
| 100GBASE-LR4, per lane | 1295, 1300, 1304, 1309 | SMF 10 km | -4.3 / +4.5 | -10.6 / +4.5 |
| 100GBASE-ER4 (ER4L), per lane | 1295, 1300, 1304, 1309 | SMF 40 km with host FEC | -2.5 / +6.5 | -20.5 / -3.5 |
The patterns in the table
Reach is bought with transmit power at one end and sensitivity at the other. Compare 10GBASE-SR at -7.3 dBm minimum launch with 10GBASE-ZR at 0 dBm, and their sensitivities at -9.9 and -24 dBm. The ZR optic launches about five times more power and hears signals about 25 times fainter. That is where 80 km comes from.
Sensitivity gets worse as the line rate goes up. On single-mode at 1310 nm, the 10G part is specified to -14.4 dBm, the 25G part to -13.3 dBm, and each lane of the 100G part to -10.6 dBm. There is less energy in each bit at a higher rate, so the receiver needs more light to pick it out of its own noise. A 10 km optic is a 10 km optic at every rate, but the budget behind it gets tighter each generation.
Long-reach optics have a high receive floor as well as a high ceiling. 10GBASE-ZR cannot accept anything stronger than -7 dBm, and 1000BASE-ZX nothing stronger than -3 dBm. Plug either into a short fibre and the link will not work until you attenuate it.
Single-fibre optics come in pairs. The 1000BASE-BX entry has two part numbers, D and U, which transmit on different wavelengths and receive on the other one. They must be deployed as a matched pair, one at each end; two of the same type facing each other will never pass light.
What is not in the table
Coarse and dense wavelength division multiplexing optics (CWDM and DWDM) each occupy a single narrow channel and are specified individually per wavelength, so there is no single row for them. They behave exactly as described in this guide, with one addition: the multiplexer and demultiplexer are themselves components with insertion loss, and that loss belongs in the budget alongside the fibre and the connectors.
Single-mode, multimode and the cost of mixing them
The two fibre families differ mainly in how wide the light-carrying region is.
Multimode fibre, as TIA-568-C.3 specifies it, is 62.5/125 µm or 50/125 µm: a core of 62.5 or 50 micrometres inside a 125 micrometre cladding. The wide core lets light travel by many paths at once, which is why it is cheap to couple into and why those paths arriving at slightly different times eventually limit the distance.
Single-mode fibre carries one path. ITU-T G.652 specifies it by mode field diameter rather than core diameter, with a nominal range of 8.6 to 9.5 µm at 1310 nm for G.652.B and 8.6 to 9.2 µm for G.652.D, in the same 125 µm cladding.
So the two have the same outside diameter and roughly a six-to-one difference in the part that matters. Every connector will mate with every other connector, and nothing in the physical fit will warn you that you have joined the wrong two things.
What happens when you mix them
Joining the two directions behaves very differently:
Single-mode into multimode is the geometrically forgiving direction. A narrow beam entering a wide core mostly gets in. The link may even come up over a short distance, which is what makes this mistake persistent: it half works, intermittently, and nobody suspects the fibre.
Multimode into single-mode is the unforgiving direction. A wide beam arriving at a narrow core loses most of its power at the joint, and the receiver sees a deficit far larger than anything the loss budget would explain.
In both directions the symptom in the readings is the same shape: Tx power normal at both ends, Rx power far below what the budget predicts, and the deficit constant over time. It does not drift, it does not respond to cleaning, and it is usually too big to be a connector.
Wavelength mismatch
The other half of the same mistake is putting two different optic types at the two ends of a link. An 850 nm short-reach module facing a 1310 nm long-reach module will not pass traffic in either direction, whatever the fibre.
This is easier to make than it sounds, because the two modules look alike, both report healthy Tx power, and both report no Rx power. Reading the module identity, not just the power, is what settles it: the vendor name, part number and wavelength all sit in A0h and are printed by the same commands.
Bidirectional single-fibre optics add a twist, since the matched pair is supposed to use two different wavelengths. Two of the same variant facing each other is the failure; one of each is correct.
Telling them apart before you plug them in
TIA-568-C.3 clause 5.2.3 gives an advisory colour code for connector strain relief and adapter housings:
| Fibre | Colour |
|---|---|
| 62.5/125 µm multimode | Beige |
| 50/125 µm multimode | Black |
| 850 nm laser-optimised 50/125 µm multimode (OM3) | Aqua |
| Single-mode | Blue |
| Single-mode, angled contact ferrule | Green |
Two cautions. The clause is advisory and applies unless the colour coding is being used for something else, so a site with its own convention is not non-compliant. And cable jacket colour is a different standard: TIA-568-C.3 defers it to ANSI/TIA-598-C.
The green strain relief is the one to recognise on sight. It marks an angled physical contact (APC) connector, whose ferrule is polished at an angle to reduce back reflection. An angled end face cannot seat properly against a flat-polished (UPC) one, so the two types are not interchangeable and should not be mated to each other. Green goes with green.
Fault patterns: reading the numbers as a diagnosis
The readings are most useful as a pair, taken at both ends. Tx tells you whether the module is working. Rx tells you whether the path is working. Together they narrow the fault to one of a small number of causes before anyone walks anywhere.
Always collect four numbers, not two: Tx and Rx at your end, Tx and Rx at the far end.
The order matters. Checking Rx first invites you to blame the fibre for what is actually a dead laser at the other end.
The same thing as a lookup table
| Local Tx | Local Rx | Far Tx | Far Rx | Most likely cause |
|---|---|---|---|---|
| Normal | Normal | Normal | Normal | Healthy. If traffic is still bad, it is not an optical power problem |
| Normal | None | Normal | None | Path broken in both directions, or patched to the wrong port |
| Normal | Normal | Normal | None | One strand broken or mis-patched, in the direction feeding the dark end |
| Normal | Low | Normal | Low | Excess path loss: dirt, bends, too many connections, or too long |
| Normal | Low | Normal | Normal | Loss on one strand only: a single bad connector or lead |
| Normal | Too high | Normal | Too high | Long-reach optic on a short link; attenuate |
| Low, bias high | Any | Normal | Low or none | Local transmitter failing |
| Zero, bias zero | Any | Normal | None | Local laser off: disabled, shut down, or module fault |
| Normal | Normal | No reading | No reading | Far end has no diagnostics, or you are not actually reading the far end of this link |
"Low" means below the data sheet sensitivity or below the module's low warning threshold, whichever you have to hand.
Three cases the tree does not cover
Rx power that moves around. A reading that swings by several decibels between polls, especially one that changes when someone is working in the room or when the building heats up, points at a mechanical fault rather than a loss budget: a connector not fully seated, a cracked ferrule, a lead pinched under a cabinet door, or a patch cord bent hard around a corner. Loss from dirt and length is constant. Loss that varies has something moving.
Rx power that declines slowly. A drift of a decibel or two over months, on a link nobody has touched, is the signature of a plant that is ageing or a connector that is gathering contamination. This is the case that justifies trending every optic: the absolute reading is only good to 3 dB, but the trend of one module against itself is reliable, and it gives you months of warning.
One lane of four. On QSFP and QSFP28, read the per-lane figures. A single weak lane will show as degraded throughput and rising errors rather than a down link, and any tool that reports one averaged power for the module will show nothing wrong at all.
Attenuators: when there is too much light
An attenuator is a component that throws away a known number of decibels. It exists because the alternative, a receiver running above its overload point, produces errors that look like everything else.
When you need one
The FOA's summary is the right rule of thumb: multimode systems generally do not need attenuators, while single-mode systems, especially short links, often have too much power and need them. The reason is in the table in section 9. A multimode optic launches at around -7 dBm and overloads at around -1 dBm, so the fibre between them rarely delivers too much. A long-reach single-mode optic launches at 0 dBm or above and overloads at -7 dBm or lower, so on a short fibre the light arrives well above what the receiver will take.
Vendor documentation says the same thing in specific terms. Cisco's gigabit SFP data sheet calls for a 5 dB inline attenuator at each end for back-to-back connections on 1000BASE-EX, and a 10 dB attenuator at each end on 1000BASE-ZX when the cable span loss is below 8 dB.
The situations that produce the problem are predictable:
A long-reach optic used on a short link because that was the spare in the cupboard.
Two devices connected back to back in a rack with a one metre patch lead.
A link that was long and got shorter, usually because a route was re-engineered.
A lab or staging build of a link that will be long in production.
Sizing one
attenuation needed = measured Rx power - target Rx powerFor the target, the FOA advises aiming near the middle of the receiver's operating range rather than near sensitivity or overload. Take the midpoint between the data sheet's maximum and minimum receive power.
A worked case. A 10GBASE-ZR link is 1 km long and the receiver reports -3 dBm. The data sheet window is -24 to -7 dBm, so the midpoint is about -15.5 dBm, and the required attenuation is -3 minus -15.5, which is 12.5 dB. Fixed attenuators come in discrete values, so pick the nearest one that lands comfortably inside the window: a 10 dB part gives about -13 dBm, which is 6 dB clear of overload and 11 dB clear of sensitivity. That will do.
Two checks before you order:
Attenuators are specified for a wavelength. A part characterised at 1550 nm does not attenuate by its marked value at 1310 nm.
Both directions need their own. The link is two fibres and the problem is usually symmetric, so you normally fit one at each end, which is exactly what the Cisco guidance above describes.
Fitting one
The FOA is specific about placement: install the attenuator at the receiver. That makes the power easy to measure, and it keeps reflections from travelling back to the transmitter. For single-mode in particular, the FOA advises choosing an attenuator type with good reflectance specifications, because some designs, gap-loss types especially, can disturb the transmitter.
After fitting, verify. The FOA's instruction is to test the system power with the transmitter on and the attenuator installed at the receiver, then confirm the reading sits within the receiver's specified range. If you are working from the module's own diagnostics rather than a meter, remember the 3 dB of uncertainty and leave yourself more headroom than the arithmetic strictly needs.
What an attenuator is not for
It does not fix a weak link. If Rx power is below sensitivity, adding loss makes it worse. The only direction an attenuator helps is down.
Worked example, end to end
Two buildings 3.5 km apart, joined by outside-plant single-mode fibre, with 10GBASE-LR optics at each end. Inside each building the cable terminates on a patch panel through a fusion-spliced pigtail, and a patch cord runs from the panel to the switch.
The addresses, distances and readings here are illustrative, but the arithmetic and the standard values are the real ones.
The budget on paper
| Item | Count | Unit loss | Loss |
|---|---|---|---|
| Outside-plant single-mode fibre | 3.5 km | 0.5 dB/km | 1.75 dB |
| Mated connector pairs | 4 | 0.75 dB | 3.00 dB |
| Fusion splices | 2 | 0.3 dB | 0.60 dB |
| Total path loss | 5.35 dB |
| Line | Value |
|---|---|
| Minimum launch power, 10GBASE-LR | -8.2 dBm |
| Less total path loss | -5.35 dB |
| Worst-case power at the receiver | -13.55 dBm |
| Receiver sensitivity, 10GBASE-LR | -14.4 dBm |
| Worst-case margin | 0.85 dB |
Less than one decibel, on a link that is a third of the optic's rated distance. That is the honest answer for a plant built exactly to the limits the cabling standard allows, and it is why 10 km on the data sheet does not mean 10 km of anything you like. Add one more patch panel and the budget goes negative.
In practice most installed plant measures better than the maxima, which is why links like this work. But the budget is what tells you how much of that good fortune you are relying on.
What the modules actually report
At commissioning:
| Reading | Building A | Building B |
|---|---|---|
| Tx power | -5.6 dBm | -5.5 dBm |
| Rx power | -9.0 dBm | -9.1 dBm |
Two things to take from this. The transmitters are launching about 2.6 dB above their specified minimum, which is normal and is where most of the real-world margin comes from. And the implied path loss, far-end Tx minus near-end Rx, is -5.6 minus -9.1, or 3.5 dB into building B, and -5.5 minus -9.0, the same 3.5 dB into building A, against the 5.35 dB the worst-case budget assumed. The plant is better than the standard requires, and it is symmetric, which is what you want to see.
Measured margin against sensitivity is -9.1 minus -14.4, or 5.3 dB. Comfortable.
Six months later
The monitoring system flags the link. New readings:
| Reading | Building A | Building B |
|---|---|---|
| Tx power | -5.6 dBm | -5.5 dBm |
| Rx power | -9.0 dBm | -13.4 dBm |
Work it through the decision tree in section 11:
Tx normal at both ends? Yes, both unchanged. Not a transmitter fault.
Any Rx power present? Yes, at both ends. Not a break.
Rx inside the window? At building A, yes. At building B, -13.4 dBm is still 1 dB above the -14.4 dBm sensitivity, but it has fallen 4.3 dB.
So one direction has lost 4.3 dB and the other has lost nothing. A change confined to one strand is not the fibre run ageing and it is not the optics, both of which would show in both directions. It is one component on the A-to-B strand: a connector that has been disturbed, contaminated, or damaged.
The response, in order: clean and reseat the connectors on that strand at both panels and both switch ports, re-read, and only if the loss persists send someone to test the strand with a light source and power meter or an OTDR. Note also that 4.3 dB is far more than the 0.75 dB a mated pair is allowed, so whatever is wrong is a fault rather than a tolerance.
One last thing this example shows. At -13.4 dBm the link is above sensitivity and will probably still be passing traffic, possibly with no errors at all. Nothing about the service would tell you this had happened. Only the trend did.
Quick reference
Numbers worth knowing by heart
| Value | Figure |
|---|---|
| 0 dBm | 1 mW |
| 3 dB | A factor of two in power |
| 10 dB | A factor of ten in power |
| -40 dBm | The lowest power the diagnostics can report, 0.1 µW |
| Plus or minus 3 dB | The accuracy of every Tx and Rx reading from a module |
| 0.75 dB | Maximum loss of a mated connector pair, TIA-568-C.3 |
| 0.3 dB | Maximum loss of a splice, TIA-568-C.3 |
| 0.5 dB/km | Maximum attenuation, outside-plant single-mode at 1310 and 1550 nm |
| 1.0 dB/km | Maximum attenuation, inside-plant single-mode at 1310 and 1550 nm |
| 3.5 dB/km | Maximum attenuation, multimode at 850 nm |
Commands
| Platform | Command |
|---|---|
| Cisco IOS XE | show interfaces transceiver detail |
| Cisco IOS XE, hardware module form | show hw-module subslot <slot/subslot> transceiver <port> status |
| Cisco IOS XE, stored thresholds | show hw-module subslot <slot/subslot> transceiver <port> idprom detail |
| Linux | ethtool -m <interface> |
A checklist for a suspect link
Read Tx and Rx at both ends. Four numbers.
Compare each Rx against the data sheet sensitivity and overload, not only against the module's own thresholds.
Subtract: far-end Tx minus near-end Rx gives the path loss. Do it both ways and compare.
Check the module identity at both ends: same wavelength, matched pair if bidirectional.
If Rx is low in both directions, suspect the path. If low in one, suspect one strand.
If Rx is high, fit an attenuator.
If Rx is fine, stop looking at power.
Glossary
| Term | Meaning |
|---|---|
| APC | Angled physical contact, a connector polished at an angle to cut back reflection. Conventionally green |
| Bias current | The current driven through the laser; rises as the laser ages or heats to hold output power steady |
| dB | A ratio between two powers; used for losses and gains |
| dBm | An absolute power level relative to 1 mW |
| DOM, DDM | Digital optical monitoring or digital diagnostic monitoring: the live readings a module reports |
| Insertion loss | The power lost by adding a component to the path |
| Launch power | Transmit power, the light the module puts into the fibre |
| Link budget | Launch power minus all path losses, compared against receiver sensitivity |
| Margin | How many decibels of loss the link could still absorb before failing |
| Modal bandwidth | The bandwidth-distance product of a multimode fibre; what sets multimode distance limits |
| OMA | Optical modulation amplitude, the power difference between a one and a zero |
| Overload | The highest receive power a receiver will tolerate before errors rise |
| Sensitivity | The lowest receive power at which a receiver meets its specified error ratio |
| SFF-8472 | The management interface specification for SFP-family modules |
| SFF-8636 | The equivalent for four-lane modules such as QSFP and QSFP28 |
| UPC | Ultra physical contact, a flat-polished connector. Never mate with APC |
References
Every figure in this guide comes from one of the documents below. Where a number is specific to a transceiver, the vendor data sheet for the part you actually hold is the authority, not this list.
Standards
SFF-8472, Specification for Management Interface for SFP+, Rev 12.4 - the SFP memory map, diagnostic units, calibration and accuracy requirements
SFF-8636, Management Interface for 4-lane Modules and Cables - the QSFP and QSFP28 equivalent
ANSI/TIA-568-C.3, Optical Fiber Cabling Components Standard - maximum cable attenuation, modal bandwidth, connector and splice loss, connector colour coding
ITU-T G.652, Characteristics of a single-mode optical fibre and cable - mode field and cladding diameters
The SFF specifications are published by SNIA; TIA standards are published by the Telecommunications Industry Association. The links above are to publicly posted copies.
Vendor data sheets
Cisco SFP Modules for Gigabit Ethernet Applications Data Sheet - 1000BASE-SX, LX/LH, EX, ZX and BX power ranges and attenuator guidance
Cisco 10GBASE SFP+ Modules Data Sheet - 10GBASE-SR, LRM, LR, ER and ZR power ranges
Cisco 25GBASE SFP28 Modules Data Sheet - 25GBASE-SR and LR power ranges
Cisco 100GBASE QSFP-100G Modules Data Sheet - per-lane power ranges for SR4, CWDM4, LR4 and ER4
Operations and practice
Cisco, Troubleshoot Transceivers in Cisco IOS XE - the commands, the threshold columns and a worked example of their values
NVIDIA Cumulus Linux, Monitoring Interfaces and Transceivers with ethtool - the
ethtool -mfield names, sample output and threshold fieldsThe Fiber Optic Association, Optical Attenuators - when attenuators are needed, where to fit them and how to choose the value
Lightwave, Maintaining average power, extinction ratio in transceivers - automatic power control, laser threshold current and why bias current rises