Does my link work? The optical budget, step by step
An optical module does not carry kilometres, it carries decibels. The whole question of a link between two sites comes down to one subtraction: what the module emits, minus what the fibre, the multiplexers and the connections take from it. This guide runs that calculation on two real cases, shows how to recover the decibels you are missing, and names the two faults of an older fibre that almost nobody checks before signing.
Estimate my cost →In practice: four figures to gather #
- The real fibre distance, not the straight line. A cable follows roads and railways: it commonly runs 1.3 to 1.5 times the direct distance. Your landlord knows it.
- The fibre reference. A declared G.652.D gives you two extra channels and removes a dispersion risk. Silence costs you both.
- The number of patch panels and splices crossed. Each connection costs half a decibel, each splice a tenth. Over a route with twenty of them, that is the difference between one module class and the next.
- The optical budget of your modules, read on the datasheet and not on the product name. A « 80 kilometre » module carries decibels, not kilometres.
The water peak, and the fibre you lease #
Two CWDM channels, 1371 and 1391 nanometres, land in a region where older fibre absorbs heavily, because of hydroxyl ions trapped in the glass around 1383 nanometres. On recent low water peak fibre the problem disappears. So you do not have 18 channels but 16, unless your lessor declares G.652.C or D fibre. No standardised figure exists to quantify that penalty on older fibre: only a route measurement gives it. Ask for the fibre reference before sizing anything: the answer is one line long and it changes how many channels you have.
The optical budget, the only question that decides #
A module does not carry kilometres, it carries decibels. The calculation fits on one line: the module budget, minus fibre attenuation times distance, minus the two multiplexer crossings, minus connectors and splices. What is left is your margin, and three decibels of it must be kept for ageing and future repairs. Below that threshold the link works on acceptance day and calls you back two winters later. The values used here are planning values, more conservative than datasheets, and they are shown with their origin.
Worked case: two data centers 30 kilometres apart #
Eight 10 gigabit links and one 100 gigabit link, a declared G.652.D fibre pair, two channels kept in reserve. The 100 gigabit rules out CWDM, so DWDM with 40 channels. For the 10 gigabit links the arithmetic gives 19.3 decibels of loss: 7.5 for the fibre, 9 for the two multiplexers, 2 for four connectors and 0.8 for eight estimated splices. A 40 kilometre class module would not do. An 80 kilometre class module carries 23 decibels, leaving 3.7 decibels of margin: the link works.
The 100 gigabit link, and the route nobody looks at #
Three options present themselves, and the third is nearly always right at this distance. A PAM4 modulated DWDM module requires an amplifier and dispersion compensation, disproportionate for one service over 30 kilometres. A coherent module carries only about ten decibels of budget and will not reach without amplification. A native ER4 type module, however, uses the multiplexer's 1310 nanometre port, alongside the C band channels without disturbing them: 15.3 decibels of loss, a 18 decibel budget with forward error correction enabled, and 2.7 decibels of margin. That is below the three decibel threshold, so the link works with little margin, and this guide says so rather than rounding up.
And at 60 kilometres? 4.5 decibels are missing #
- Losses climb to 27.5 decibels against 23 available. No configuration recovers that, it is physics.
- First remedy: a multiplexer with lower insertion loss, roughly three decibels regained across both ends, to be confirmed on the datasheet of the equipment actually chosen.
- Second remedy: a reflectometer measurement of the route. If the fibre is better than the planning value, that is three more decibels. It requires a measurement, not an assumption.
- Both together return 6 decibels and bring the margin back to 1.5 decibel, so a link that works by a hair, subject to two real checks.
- Third remedy: a booster amplifier at the head of the link. It exists only in the C band, so it changes the architecture and rules out the 1310 nanometre port. CWDM cannot be amplified at all: you have to move to DWDM.
- Fourth avenue, often overlooked: remove joints. Every intermediate patch panel costs half a decibel per pass.
Calculate your own link #
The calculation above is one specific case. Here is the same engine, with your figures. It gives the verdict, the bill of materials, and above all the list of what it had to assume for lack of a measurement. The result lives in the page address: you can paste it into a ticket or send it to your supplier.
Receiver overload, the short link trap #
The symmetric fault exists and is less well known: above a certain level the receiver saturates and the link turns erratic while everything looks generously sized. There is good news, though, and it is counter-intuitive: with a multiplexer at each end, insertion loss already acts as an attenuator, and saturation almost never happens. It appears when the route is both short and measured at low loss, precisely when you think you are safe. The cure is a calibrated fixed attenuator, never another module.
An older fibre has two faults, and the second one is the harsher #
The water peak costs you two channels. The second fault can cost you the whole link: polarisation mode dispersion. A fibre core is never perfectly round, and the older it is the less so. Light travels in it on two perpendicular polarisations which, in an oval core, do not travel at the same speed: the pulse spreads and ends up bleeding into the next one. The spreading grows as the square root of distance: doubling the length does not double the problem, it multiplies it by 1.4. On the other side, the budget is about a tenth of one symbol period, that is 10 picoseconds on a 10 gigabit link. Here is what that gives at 40 kilometres, by age of fibre. On recent cabled fibre, 0.32 picosecond: no issue. At the ceiling the Recommendation sets, 3.2 picoseconds: still no issue. On fibre from the 1990s, 12.6 picoseconds: the link does not work, and attenuation has nothing to do with it. Between best and worst case the ratio is forty. And the coefficient of the fibre you lease is published nowhere. It is measured, with a different instrument from the usual reflectometer. What it decides for you. If your landlord declares G.652.D fibre, the Recommendation already bounds that coefficient and the matter is closed up to 200 kilometres at 10 gigabit. If they declare nothing and the fibre is as old as the network carrying it, then past thirty kilometres or so you are gambling. Ask for the measurement, or plan a load test before committing. The calculator shows the three hypotheses side by side rather than picking one for you.
Two families of single-mode fibre, which are not sized the same way #
The fibre almost everyone lays is G.652, whose chromatic dispersion is about 18 picoseconds per nanometre per kilometre at 1550 nanometres. But a second standardised family exists, G.655, non-zero dispersion-shifted, designed for long distances: its dispersion runs from 4 to 8 depending on the maker, a quarter to a half. Sizing a regional route laid in G.655 with G.652 figures therefore overstates dispersion by a factor of two to four.
What that changes in practice: an 80 kilometre class 10 gigabit link tolerates 1600 picoseconds per nanometre. On G.652 that limit is reached around 89 kilometres. On G.655 at 8 it moves past 200. The form below accepts both families, and takes for G.655 the highest value of its range: a missing warning costs more than one warning too many.
One trade-off to know about. That low dispersion, which keeps the signal clean, also encourages four-wave mixing between closely spaced, high-power DWDM channels. If you plan forty channels at full power over G.655, put the question to your integrator.
What distance costs in time #
Latency is often the real reason a direct link gets built, and it is easy to work out. In fibre, light does not travel at 300,000 kilometres per second but at about 204,000, because glass has a group index near 1.47. That gives 4.9 microseconds per kilometre one way, so a little under 10 microseconds per kilometre round trip.
Mind what you multiply. The distance to use here is the actual fibre route, not the straight line: a cable follows roads and railways, and commonly runs 1.3 to 1.5 times the direct distance. Over 40 kilometres of fibre the physical floor is therefore 0.4 millisecond round trip, and no equipment will go below it. Multiplexing itself adds almost nothing: a passive multiplexer is a piece of glass, it costs a few nanoseconds. That is in fact the decisive argument against a carrier-operated link, which crosses active equipment whose transit time varies from moment to moment.
What it decides for you. If your need is synchronous replication between two storage arrays, the latency budget is counted in fractions of a millisecond, and it sets a maximum distance long before the optical budget becomes a problem. Check that constraint first: it rules out sites the optical calculation would have accepted.
What this changes for the dark fibre you lease #
The fibre rent is often the cheapest part of the file. What costs money is its ends: the optical modules, the multiplexers, and amplification if distance demands it. Before signing a lease, ask three things in writing, because they decide your optical budget: the fibre reference, the measured attenuation per kilometre on the real route, and the number of patch panels and splices crossed. A serious lessor answers all three, and the difference between an answer and a silence can be several decibels, therefore a module class.
What this guide does not cover #
The choice between CWDM and DWDM, how many channels are available and which modules fit in each: that is the subject of CWDM or DWDM: which system, how many channels, which modules, to read before this one if the question is still open. Also out of scope: active transponder systems, remotely reconfigurable multiplexers, the L band and submarine links.
Method, and how to contradict us #
The calculation shown here is produced by the site's engine, version 2026.09-v2, from a reference file where every value carries its source, its date and its status. Three statuses: the text of a Recommendation, a manufacturer datasheet at a given date, or a planning value more conservative than the datasheet. A value that is not published stays empty and says so, it is never filled in with a plausible estimate. This is a calculation, not a guarantee of service: a route measurement and a load test are what count. If you measure something other than what we calculate, write to us, the method is published precisely for that. Written on 13 September 2026. Submitted the same day to an external technical review, from which three corrections came: the distinction between 100 gigabit on a single wavelength and a two-subcarrier module, the width required of the 1310 nanometre port, and the real reach of 400ZR over an amplified line. Each is flagged in the text where it applies.
FAQ #
No. The common amplifier covers only the C band, and CWDM channels are spread too widely across the spectrum to fit. The reach of a CWDM link is the reach of its modules, full stop. If amplification is needed, you have to move to DWDM.
Two channels fall in a region where older fibre absorbs heavily, due to impurities around 1383 nanometres. On recent low water peak fibre they are usable. Ask your lessor for the fibre reference.
It carries a budget in decibels, not a distance. Count the fibre, both multiplexers, the connectors, the splices and the ageing margin. At 60 kilometres with two multiplexers, an 80 kilometre class module already falls short.
Rarely when there is a multiplexer at each end, because their insertion loss is enough. The risk appears on a short route measured at low loss. Compare the received power reported by the module with the saturation threshold on its datasheet.
No, and that matters. It rests on route assumptions nobody has measured, shown with their origin. A reflectometer measurement and a load test are what count. The calculation is there to choose the right equipment and avoid the bad surprise, not to replace acceptance testing.
Written on 13 September 2026.
From reading to comparing: relevant data centers
Estimate my cost → Compare data centers