Home › Guides › CWDM et DWDM

CWDM or DWDM: which system to choose, and which modules to put in it

A leased fibre pair between two sites looks narrow: two strands, one service. Wavelength division multiplexing puts 8, 18, 40 or 80 independent channels through it, each with its own optical module, with nothing active between the two buildings. What remains is to choose which of the two systems suits your number of services and their rate, and which modules fit in its channels. That is what this guide answers. The question of distance, which is measured in decibels, has a guide of its own.

Estimate my cost →

The idea, in one picture #

Several signals come in, each on a different colour. A multiplexer combines them into a single fibre. At the far end a demultiplexer separates them and hands each one back its port. Both boxes are entirely passive: no power, no configuration, no software failure. All the intelligence sits in the optical modules plugged into your equipment, and that is where the money goes. Wavelength-division multiplexing does not manufacture capacity: it shares a fibre you are already paying for.

Four services, four colours, one fibre between the two sites.

In practice: four moves before you order #

Count your services and their rate. That number decides between the two systems: up to sixteen or eighteen channels and no 100 gigabit, CWDM is enough and costs three times less. Beyond that, or as soon as a 100G appears, it is DWDM.
Read the channel width on the multiplexer datasheet, under « channel bandwidth ». It is written in nanometres or in gigahertz. That is the doorway your signal has to go through.
Compare it with the symbol rate of your modules, not with their gigabit rate. A 10G takes 10.3 gigabaud, a 100G on a single wavelength takes 53. If the window is narrower than that figure, nothing will pass, whatever power you inject.
Ask for the exact fibre reference from whoever leases it to you. A declared G.652.D gives you eighteen channels instead of sixteen. And keep that answer: it also feeds the optical budget calculation, which decides the distance.

Two grids: twenty nanometres apart, or eight tenths #

CWDM follows the grid of ITU-T Recommendation G.694.2: 18 wavelengths from 1271 to 1611 nanometres, spaced 20 nanometres apart. That generous spacing exists for a precise reason: the lasers are uncooled, they drift by several nanometres with temperature, and each channel needs room around it. DWDM follows ITU-T Recommendation G.694.1, anchored on 193.1 terahertz, with 100 gigahertz spacing, that is 0.8 nanometre, for 40 channels, or 50 gigahertz for 80. Packing channels that tightly requires stabilising the wavelength, therefore cooling the laser. The entire price gap between the two families lives there.

Above, the 18 CWDM channels across 340 nm. Below, the 40 DWDM channels that fit within 32 nm.

Channel bandwidth: the datasheet line to read #

Start with the picture, it saves a lot of confusion: a channel is a doorway, and a signal is a piece of furniture. The doorway has a width, printed on the multiplexer datasheet under « channel bandwidth », as plus or minus 0.11 nanometre or plus or minus 31.5 gigahertz. The furniture has a width too. And pushing harder does not get an oversized wardrobe through a narrow door: that is why no amplifier and no extra decibel rescues a signal that overflows its window. What gives the signal its width is where intuition fails. Not the bit rate, the symbol rate, in gigabaud. A symbol is a state the laser holds for an instant; depending on the modulation it carries a different number of bits. In PAM4 a symbol is two bits, so 100 gigabit on a single wavelength transmits at 53.1 gigabaud. In DP-16QAM a symbol is eight bits, so a coherent 400 gigabit transmits at only 59.8 gigabaud, for four times the data. The 400G takes almost the same room in the spectrum as the 100G, and modulation is what makes that possible. What it decides for you. Before ordering a multiplexer, read that line and compare it with the symbol rate of your modules. Below a ratio of 1.3 you are tight, below 1 the signal is clipped whatever you do next. It is a thirty-second check that saves replacing a unit.

Both windows and the four signals, at the same scale. The common 100 GHz multiplexer clips 100G and 400G.

Why 25G, 100G and 400G do not behave the same way #

Three market rules look like arbitrary conventions until you put the window next to the symbol rate. Why 25 gigabit is not sold over DWDM on the 100 gigahertz grid: 25.8 gigabaud in a 27.5 gigahertz window, the signal fills it completely and the slightest temperature drift clips its edges. Why a 100 gigabit PAM4 on a single wavelength does not cross an ordinary multiplexer: 53.1 gigabaud in 27.5 gigahertz, it is physically clipped. Vendors do not hide it, they word it differently: their datasheets list 1G, 10G and 25G Ethernet support on the channels, and send 40G, 100G and 400G to the 1310 nanometre port. Why the OIF defined 400ZR on a 75 gigahertz grid rather than 50: at 59.8 gigabaud, a 50 gigahertz window would have been too narrow. The standard merely records the physics. One caveat though: not every 100 gigabit PAM4 module emits a single 53.1 gigabaud signal. Some DWDM transponders spread the 100 gigabit over two subcarriers at roughly half the symbol rate each, and then take two adjacent channels instead of one. The window rule still holds, but it applies to each subcarrier rather than to the module as a whole. That per-subcarrier rate is published on no accessible datasheet: it is a question to put to the vendor before ordering.

The unit that carries 100G on every channel, and what it costs #

It exists, and it comes down to filter technology. Common multiplexers use a so-called Gaussian arrayed waveguide grating: its window is narrow and peaked. The flat-top models flatten the top of the curve and widen the window, at the price of higher insertion loss. The most accessible member of that family in reseller catalogues is a 64 channel unit over the extended C band, 75 gigahertz spacing, whose datasheet states a channel bandwidth of plus or minus 31.5 gigahertz, that is a 63 gigahertz window. Enough for a 100G PAM4 at 53.1 gigabaud, and just barely enough for a 400ZR at 59.8. But read the whole datasheet before buying, because it is paid for three times over.

	Gaussian AAWG, 100 GHz	Flat-top AAWG, 75 GHz
Channels	40	64
Window per channel	27.5 GHz	63 GHz
Max rate per channel	25G, already at the limit	400G, at the limit
Typical insertion loss	3.5 dB	5.8 dB
Over the link, two units	7 dB	11.6 dB, that is 4.6 dB more
Adjacent channel isolation	at least 25 dB	at least 8 dB
1310 nm port	yes	no

Which module to put in front of a given window #

The rule also reads backwards, and that is how you use it in front of a supplier datasheet. Read the channel bandwidth, convert it to gigahertz if it is given in nanometres, and you immediately know which modules stand a chance. The conversion is one line: width in gigahertz equals 299,792 times the width in nanometres, divided by the square of the wavelength in nanometres. At 1550 nanometres, 0.22 nanometre is 27.5 gigahertz. Then a signal must never be wider than its window, and it had better use less than three quarters of it.

Module	Symbol rate	27.5 GHz window	63 GHz window	CWDM window
SFP 1G	1.25 GBd	fits	fits	fits
SFP+ 10G	10.3 GBd	fits	fits	fits
SFP28 25G	25.8 GBd	fills the window	fits	fits
QSFP28 100G PAM4	53.1 GBd	clipped	passes	no such module
QSFP-DD 400G ZR	59.8 GBd	clipped	fills the window	no such module
QSFP+ 40G LR4	4 internal lanes	1310 nm port	no 1310 port	1310 nm port
QSFP28 100G ER4	4 internal lanes	1310 nm port	no 1310 port	1310 nm port

« I need 40 gigabit between A and B » #

A common request, and the right answer is almost never a 40 gigabit module. Three routes exist. Four aggregated 10 gigabit links, over four wavelengths: cheapest modules on the market, the optical budget of an ordinary 10G, and this is the answer in the vast majority of cases. One caveat, a single flow will never exceed 10 gigabit: awkward for storage array replication, irrelevant for a site uplink. One 40GBASE-LR4 module through the 1310 nanometre port: it carries four internal lanes in the O band, so it uses no C band channel at all, but there is only one 1310 port per unit, and only one service can take it. Watch one detail almost no datasheet publishes: that module spreads its four lanes from 1295.56 to 1309.14 nanometres, so the port must pass the whole range. A port cut narrowly around 1310 nanometres would clip the lower lanes, and the link would fail with the optical budget entirely innocent. Ask for the exact port range before ordering. Move straight to 100 gigabit: 40G is an abandoned generation, its modules often cost more than their 100G equivalents, and one wavelength is enough. The calculator below now accepts 40 gigabit and returns all three routes, with the optical budget of the one it recommends.

The combinations that exist, and those that do not #

1 gigabit: in CWDM as in DWDM, across the whole grid, up to around 80 kilometres.
10 gigabits: everywhere too. In CWDM, the upper band from 1471 to 1611 nanometres has 80 kilometre modules, the lower band stops at 40.
25 gigabits: in CWDM only in the O band, around 1271 to 1331 nanometres, over short distances. The optical budgets of those modules are not published in the datasheets we could read, so we do not calculate that line.
100 gigabits: never on a CWDM wavelength. Three routes in DWDM, and a fourth that often saves the day, the multiplexer's 1310 nanometre port with a native ER4 type module.
400 gigabits: coherent DWDM only. Read the datasheets carefully: the advertised 40 kilometres is point to point WITHOUT a multiplexer, because those modules carry only about ten decibels of budget.

And now, does it work? #

You know which system to choose, how many channels you have and which modules fit in them. The second question remains, and it is the one that really decides: does the light reach the other end? It depends on distance, on the age of the fibre and on the number of connections, and it can be calculated. It has its own guide, with the calculator: Does my link work? The optical budget, step by step. You will find the detailed calculation on two real cases, at 30 and at 60 kilometres, the costed remedies when decibels are missing, and the two faults of an older fibre that almost nobody checks before signing.

Which multiplexer to buy, by type #

Over a fibre pair: the basic form, one box per end, N channels for N duplex services.
Over a single strand, called bidirectional: both directions share the fibre, so every service consumes two wavelengths. Your capacity halves, and the boxes come in pairs with mirrored assignments.
Add and drop, for an intermediate site or a ring: only that site's channels are extracted, the others pass through with a small loss that must be added on the channel travelling furthest.
The ports that change everything: the expansion port to chain a second multiplexer, the monitor port to attach a reflectometer without cutting service, and the 1310 nanometre port to run an O band service alongside the C band channels.
What to read on a datasheet: maximum insertion loss rather than typical, isolation between adjacent channels, and the passband width per channel.

Migrating from CWDM to DWDM without touching the fibre #

This is the cheapest upgrade path, and the equipment is built for it. Some CWDM multiplexers offer a 1550 nanometre port matching one of the grid channels. A complete DWDM system plugs in there and occupies that single CWDM channel as far as the first system is concerned. You keep your existing CWDM services, you add 40 channels, and you migrate gradually. No work on the fibre itself, which matters when it is leased and every intervention has to be negotiated.

What this guide does not cover #

Active transponder systems and remotely reconfigurable multiplexers, which belong to another trade and another budget. The L band, used when the C band is full. Submarine links. And buying a wavelength from a carrier rather than building your own system, which is often the right answer when you have one or two services and no optical team. This guide covers the common case: a leased fibre pair between two sites, to be lit yourself.

Method, and how to contradict us #

The grids quoted here are those of the ITU-T Recommendations, the equipment characteristics come from public datasheets whose reading date is given, and the symbol rates from IEEE 802.3 and OIF standards. A value that is not published stays empty and says so: it is never filled in with a plausible estimate. Written on 13 September 2026. If you measure something different from what we publish, write to us.

FAQ #

CWDM or DWDM, how do I choose in one sentence?

Up to eight services at 10 gigabits maximum and 80 kilometres, CWDM is enough and costs considerably less. Beyond that in channel count, or as soon as a single service runs at 100 gigabits, it is DWDM.

Can I do 100 gigabits over CWDM?

Not on a CWDM wavelength. Modules labelled CWDM4 carry four wavelengths inside the module, and a passive multiplexer can neither take them on one port nor separate them. The simple route at short distance is the DWDM multiplexer's 1310 nanometre port with a native module.

How many services on a single strand rather than a pair?

Half, because each service consumes one wavelength per direction. An 18 channel CWDM system carries only 9 services, and the boxes are bought in pairs with mirrored assignments.

What is an add and drop multiplexer for?

To serve an intermediate site without breaking the main link. Only that site's channels are dropped and added, the others pass through with a small loss. It is what lets you cable three data centers in a ring with a single system.

Written on 1 September 2026.