An optical module is not a radio. Its hardest electrical requirement is a 224 Gb/s PAM4 signal whose energy sits around 56 GHz, carried a few centimetres from a driver to a modulator, inside a package small enough that the laminate also has to behave mechanically like a substrate. That combination — ultra-low loss, micrometre-scale dimensional stability, plated microvias and eight reflows — is what makes optical module laminates a separate list from ordinary high-frequency materials. This article summarises the high frequency laminate models in use across the 400G, 800G and 1.6T generations, with the loss arithmetic that decides between them and the requirement most often left out of a comparison table: dimensional stability.
What an Optical Module Asks of a Laminate
Five requirements apply at once, and they pull in different directions in a way that ordinary high-frequency work does not.
The electrical requirement is the obvious one: an insertion-loss budget measured in tenths of a dB over a channel a few centimetres long. The mechanical requirement is the one that decides yield: the laser, the modulator and the detector have to be aligned to their driver electronics to within a few micrometres, through several reflow cycles, on a laminate that is also being asked to behave like a package substrate. The process requirement is HDI with laser-drilled microvias, because the package cannot be bigger than it is. The thermal requirement is a Tg and Td high enough to survive assembly without smear or delamination. And the supply requirement in the current market is a second source, which is why domestic ultra-low-loss grades now appear on the same comparison tables as the established ones.
Most published comparisons cover the first requirement and skip the second. That is a mistake in this application specifically, because a laminate that meets the Df target and moves 50 µm during assembly produces a module that fails at the optical interface rather than at the electrical one — and the electrical test passes.
Where the Energy Sits: 224 Gb/s PAM4 and the 56 GHz Nyquist
The number that organises the whole material discussion is not the data rate but the Nyquist frequency. A 224 Gb/s PAM4 lane runs at 112 GBd, so its fundamental energy sits at 56 GHz, with meaningful content out to about 1.5 times that. The previous generation, 112 Gb/s PAM4, sits at 28 GHz.
Loss scales linearly with frequency, so a laminate assessed at 10 GHz is being extrapolated by a factor of 5.6 to reach 56 GHz. That extrapolation is the single largest source of disagreement between a datasheet and a working module, and it is why the useful comparison figure is the dielectric loss at the Nyquist frequency itself rather than the Df value quoted at whatever frequency the supplier chose.
| Laminate class | Dk | Df | Dielectric loss at 28 GHz | At 56 GHz |
|---|---|---|---|---|
| Standard FR-4 | 4.15 | 0.020 | 2.98 dB/in | 5.95 dB/in |
| RO4350B / RO4835 | 3.48 | 0.0037 | 0.50 dB/in | 1.01 dB/in |
| RO4003C | 3.38 | 0.0027 | 0.38 dB/in | 0.77 dB/in |
| Ultra-low-loss digital, Df 0.002 class | 3.6 | 0.0020 | 0.27 dB/in | 0.55 dB/in |
| RO3003 | 3.00 | 0.0010 | 0.13 dB/in | 0.25 dB/in |
| RT/duroid 5880 | 2.20 | 0.0009 | 0.10 dB/in | 0.20 dB/in |
Two readings matter. The gap between a 0.0020 digital grade and a 0.0037 RF grade is 0.46 dB/in at 56 GHz — over a 30 mm channel that is 0.54 dB, which is a real fraction of a PAM4 link budget. And the gap between 0.0020 and 0.0009 is a further 0.35 dB/in, which is why the question “does this module need PTFE” comes down to channel length rather than to ambition.
Requirement 1: Df — How to Set the Threshold Instead of Copying One
The widely quoted target for 800G work is Df below 0.005 at 10 GHz, with the top tier below 0.003. Those numbers are reasonable, but they are usually copied rather than derived, and the derivation takes three lines.
Take the channel length, multiply by the dielectric loss per unit length at 56 GHz, and compare with the loss budget left after the connector, the package and the DSP equalisation. A 30 mm (1.2 in) channel with 0.5 dB of budget therefore allows 0.42 dB/in of dielectric loss, which corresponds to a Df of about 0.0015 at Dk 3.6 through αd ≈ 2.6·f·√Dk·Df. A 15 mm channel doubles the allowable Df. That is why two module makers with the same data rate can legitimately specify different materials, and why a single industry threshold is a starting point rather than an answer.
| Channel length | Dielectric loss allowed by a 0.5 dB budget | Implied Df at Dk 3.6, 56 GHz | Which class that requires |
|---|---|---|---|
| 10 mm | 1.27 dB/in | 0.0046 | RO4350B or RO4835 are adequate |
| 20 mm | 0.64 dB/in | 0.0023 | RO4003C at the limit; 0.002 class comfortable |
| 30 mm | 0.42 dB/in | 0.0015 | 0.002 class with margin, or PTFE |
| 50 mm | 0.25 dB/in | 0.0009 | PTFE only — 5880 or TLY-5 territory |
The table is why the same data rate produces different material choices in different module formats. A short-reach module with a 10 mm channel is well served by a hydrocarbon; a 50 mm channel in the same generation is not, and no amount of equalisation recovers 0.5 dB of pure dielectric loss without paying for it in power and latency.
Conductor loss has to be added, and at 56 GHz it is no longer a minor term. On a low-Df laminate the two contributions are comparable, which is why foil roughness moves from a footnote to a design parameter in the next section.
Requirement 2: Dimensional Stability, the Requirement That Decides Yield
Dimensional stability is quoted as a percentage change in length after a defined thermal excursion, and 0.05 % is the usual requirement for optical module work. On a 100 mm span that is 50 µm, which is one to two orders of magnitude larger than the alignment tolerance between a laser and its coupling lens.
What this means in practice is that the alignment cannot be established before the last reflow. The module is designed so that optical elements are actively aligned and fixed after the board has been through its thermal history, or so that the board is dimensionally stable enough that the movement is repeatable rather than random. Both routes need the laminate to be isotropic and to have matched in-plane CTE: a hydrocarbon grade at 10 ppm/°C in x and 12 ppm in y will move differently along the two axes, and a layout that assumes symmetry will be wrong along one of them.
The requirement also rules out some otherwise excellent stacks. A hybrid that puts PTFE under the optical engine and FR-4 elsewhere has an in-plane CTE step of 30 ppm/°C or more at the boundary, which is exactly where the alignment-critical features sit. In an optical module the hybrid argument that works for a radar board works less well here: the area is small, the layer count is low, and the mechanical penalty is paid in the most sensitive region of the product.
Requirement 3: Copper Foil, Because the Skin Is 0.28 Micrometres Deep
At 56 GHz the skin depth in copper is 0.28 µm, calculated as 2.1 µm divided by the square root of the frequency in gigahertz. Current therefore flows in a layer thinner than the surface profile of a standard foil, which means the current path is longer than the geometric length of the trace and the conductor loss rises accordingly.
The practical consequence is that ultra-low-profile and HVLP foils are not optional on a 56 GHz channel. The difference between a standard foil and an HVLP foil in conductor loss at this frequency is a few tenths of a dB per inch, which is the same order as the difference between two adjacent Df classes — so a material comparison that holds Df constant and ignores foil will rank the grades wrongly.
Two further consequences follow. Surface treatment has to be chosen for loss rather than for solderability: at 56 GHz the current is entirely in the finish, and a nickel-bearing finish such as ENIG puts a higher-resistivity layer exactly where the current flows. And roughness has to be specified per side, because the foil side facing the dielectric and the side facing the plating do not have the same profile.
Requirement 4: HDI, Microvias and the Stack That Fits in a Module
An optical module is built on a substrate-like PCB: 4 to 8 layers, microvias of 75–100 µm, via-in-pad, and line widths at the limits of what a panel process can hold. The rules from mmWave work apply with one difference — the driver is density rather than wavelength.
The aspect-ratio rule is the same: a laser-drilled microvia is reliable at a depth-to-diameter ratio of about 1:1, with 0.75:1 preferred, so a 100 µm via requires a dielectric of 100 µm or less. That is compatible with a 4-layer module core and incompatible with the 0.254 mm dielectric that a 50 Ω RF line would prefer, which is the tension the stack has to resolve. In modules the resolution is usually a thinner dielectric and a narrower line, accepting higher conductor loss in exchange for the density the package needs.
Via-in-pad also forces a resin-plug or copper-filled via process, and both have a thermal cost: the filled via becomes a mechanical feature in a board that is being asked for 0.05 % stability, and the fill material’s CTE is not the laminate’s. This is one of the places where a low-CTE laminate helps twice — once in the dielectric and once in what it does to the via fill.
Requirement 5: Thermal History and CAF
An optical module sees multiple reflows: the transceiver assembly, the module-to-board attach, and often a rework pass. Each one is a thermal excursion that the laminate has to survive without smear, delamination or a change in dimensional stability, and the cumulative effect is what Td and Tg are proxies for.
The numbers worth checking are Td rather than Tg for this application: RO4003C decomposes at 425 °C and RO4350B at 390 °C, while the ceramic-filled PTFE grades in the 3000 series are quoted at 500 °C. For the ultra-low-loss digital grades, high Tg and CAF resistance are usually quoted together, because the same resin chemistry that raises Tg also reduces the ionic mobility that CAF needs.
CAF is a real qualification item here rather than a datasheet courtesy. A module sits in a humidity-controlled but not humidity-free environment for a decade, at a bias of a few volts and a pitch of tens of micrometres between features. That is the exact combination that grows conductive filaments, which is why the CAF test result belongs on the comparison table next to Df and why a grade with excellent loss and a mediocre CAF result is not a candidate.
The Models: A Summary by Family
The table below summarises the high frequency laminate models that appear in optical module designs, grouped by what they are rather than by who makes them. The Df classes are the published figures at or near 10 GHz; where a supplier quotes at another frequency the class should be re-checked rather than assumed.
| Family | Representative models | Df class at 10 GHz | Where it fits in a module | Process note |
|---|---|---|---|---|
| Ultra-low-loss digital (modified epoxy / PPO) | Panasonic MEGTRON 7 and 8, Taiyo TU-872 class, Shengyi S7439 / S8 class, Isola I-Tera class | 0.002 and below | Driver-to-DSP digital section, the main signal path | Standard FR-4 process, high Tg, CAF data available |
| Low-loss hydrocarbon / ceramic | RO4350B, RO4835, RO4003C | 0.0027–0.0037 | RF driver section, flex-to-board transition, TOSA/ROSA feed | Thermoset, presses at 190–200 °C, no PTFE chemistry needed |
| PTFE / ceramic, non-woven | RT/duroid 5880, RO3003, Taconic TLY-5 | 0.0009–0.0010 | Only where the channel length justifies it; RF test structures | Sodium-naphthalene or plasma desmear, 260–300 °C press, poor fit for fine HDI |
| High-speed, high-Tg FR-4 | Mid-loss digital grades, Tg 170–180 | 0.008–0.012 | Power, control, low-speed housekeeping | Cheapest option, adequate outside the signal path |
The rankings behind that table are consistent across suppliers even where the numbers differ slightly: the PTFE grades are four times lower loss than the best digital grades, and the best digital grades are two to three times lower than the RF hydrocarbons. What changes between generations is where the line is drawn. At 400G a 0.0037 hydrocarbon was adequate for the whole module; at 800G the digital path moved to the 0.002 class; at 1.6T the question is whether the RF section moves to PTFE at all or whether a 0.002 digital grade with an improved foil is enough.
For the grade-by-grade figures behind this table, the Rogers material guide and the Taconic laminate performance summary are the two references worth keeping open, and the broader field is surveyed in the types of high-frequency and high-speed materials.
Where an RF Material Is the Wrong Answer
There is a persistent assumption that the lowest-loss laminate is always the best choice for a high-speed channel. In an optical module it is frequently wrong, for three reasons that have nothing to do with loss.
Process. A PTFE grade cannot be laser drilled with the same parameter set, cannot be desmeared with permanganate, and does not press in the same cycle as the rest of the board. A module that needs 75 µm microvias in a 4-layer HDI stack is asking the most difficult process on the most difficult material.
Mechanics. 5880 expands 237 ppm/°C in the z axis against 32 ppm for a hydrocarbon, on a board that has been asked for 0.05 % dimensional stability in plane. The via fill, the plated barrel and the optical alignment all sit inside that mismatch.
Diminishing return. The 0.35 dB/in advantage of PTFE over a 0.002 digital grade only pays if the channel is long. Inside a module the channel is centimetres, and the difference is tenths of a dB, which the DSP equaliser absorbs. The same material in a 300 mm backplane is a different calculation.
This is why the practical answer for most 800G modules is a 0.002-class digital laminate for the signal path and a hydrocarbon for the RF driver section, with PTFE appearing only where a specific channel or a test structure demands it. The comparison of hydrocarbon grades in RO4350B and RO4835 covers the RF side of that split; the ultra-low-loss case for PTFE is set out in the 5880 analysis.
Hybrid Stacks in a Module: When They Help and When They Hurt
A hybrid stack is the obvious way to combine a 0.002 digital laminate with a low-loss RF section, and it works when the boundary is a connector or a defined interface. It fails when the boundary runs through the alignment-critical area, and it fails quietly.
Three rules make it survivable. Keep the RF section contiguous — one material from the driver pad to the optical package interface, so that no RF line crosses a Dk step. Keep the boundary mechanically simple: a straight line parallel to a board edge, not a region that surrounds the optical engine. And press in two cycles with the interface layer bonded first, because the alternative — a low-flow prepreg chosen for its melting point — puts the RF bond strength into a material selected for the wrong reason.
Where the module is small enough, the simplest answer is often to make the whole board the 0.002 class and skip the hybrid. The cost difference against a two-material stack is smaller than the process risk, and the dimensional-stability requirement stops being a boundary condition.
What a Comparison Table Usually Leaves Out
Four items are missing from most laminate comparisons for this application, and each has decided a build.
The frequency at which Df was measured. A 0.0020 at 1 MHz and a 0.0020 at 10 GHz are not comparable claims, and at 56 GHz neither is directly the working number.
Dimensional stability test conditions. The percentage means nothing without the thermal excursion that produced it. A grade quoted at 0.05 % after one reflow and one quoted after three are different materials for this purpose.
Foil and finish. At 56 GHz the total loss depends on the foil profile and the surface finish as much as on Df. A comparison that fixes Df and varies foil will produce a ranking that does not reproduce on a real board.
Lot consistency. A module is built in volumes where several laminate lots pass through one line, and a Df spread of 0.0005 between lots at 56 GHz is 0.14 dB/in — more than the difference between two adjacent grades. This is the item to ask for as data, and the reasoning is the same as for any selection on key indicators: the number that matters is the one that varies, not the one that is specified.
The competitive structure behind these grades, and why the M7/M8 class is difficult to displace, is analysed in the technical barrier analysis for Panasonic M7/M8.
A Selection Checklist for a Module Build
| Question | Accept | Why it decides |
|---|---|---|
| Df at the Nyquist frequency, not at 1 MHz? | ≤0.002 for a 224 Gb/s lane on a long channel | Loss is linear in frequency; 56 GHz is 5.6× the 10 GHz figure |
| Dimensional stability after three reflows? | ≤0.05 % | Optical alignment tolerance is micrometres on a 100 mm span |
| Foil profile and finish specified? | HVLP or ULP, finish chosen for loss | Skin depth is 0.28 µm at 56 GHz |
| Microvia size and aspect ratio? | ≤100 µm at AR ≤1:1 | Dielectric must be no thicker than the via diameter |
| CAF data at the operating bias and pitch? | Tested, not estimated | A decade at a few volts and tens of micrometres |
| Second source qualified? | Same Df class, same press window, same foil | Supply continuity without requalifying the stack |
XCE PCB builds these stacks — hybrid PTFE, hydrocarbon and ultra-low-loss digital laminates in the same shop: 2 to 72 layers, panels to 600×1200 mm, impedance to ±10 %, samples in as fast as 24 hours, inspected to IPC-A-600G and IPC-6015. For optical module work the coupon is the deliverable that matters, because the number that has to be held is insertion loss at 56 GHz rather than impedance alone.
Conclusion
The high frequency laminate models used in optical modules fall into three useful classes rather than the dozen product names they are sold under. Digital grades at Df 0.002 and below carry the 224 Gb/s signal path at 0.55 dB/in at 56 GHz; hydrocarbon and ceramic grades at 0.0027–0.0037 carry the RF driver section at 0.77–1.01 dB/in; and PTFE grades at 0.0009–0.0010 reach 0.20 dB/in where the channel length justifies it, at the cost of a 260–300 °C press cycle and a 237 ppm/°C z-axis expansion that a 0.05 % stability requirement does not welcome.
The decision is therefore not “which laminate has the lowest Df” but “what channel length and what process window does this module allow”. Answered that way, most 800G designs land on a 0.002-class digital laminate with a hydrocarbon RF section, and the lowest-loss materials stay where they belong: in the test structures, the long backplane runs and the applications where a tenth of a dB decides the product.
Frequently Asked Questions
What Df does a 224 Gb/s PAM4 optical module need?
Why is dimensional stability so important in an optical module?
Is PTFE the best choice for a 1.6T module?
What copper foil should a 56 GHz channel use?
Does a lower Dk help at 56 GHz?
How many reflows does an optical module laminate have to survive?
Selecting a Laminate for a 56 GHz Channel?
Send us your Gerber files and requirements for a manufacturability review and a tailored quotation.
sales1@xcepcb.com
+86 180 3341 4657