Ten Critical Parameters for High-Frequency PCBs: A Comprehensive Guide from Dielectric Constant to Surface Roughness

The ten critical parameters for high-frequency PCBs are usually presented as a list of equals, which is why they are so hard to use. They are not equals. Three of them decide loss, two decide phase and delay, one is not a parameter at all but the result of the others, and three decide whether the board survives its own assembly. Sorting them that way turns a glossary into a decision tool: for a given link, two or three of the ten matter and the rest are background. This guide gives each one a target number, the mechanism that makes it matter, and the size of the mistake when it drifts.

Ten Critical Parameters for High-Frequency PCBs

Ten Parameters, Four Jobs

The list below is the same ten items found in every textbook, but grouped by what they do to a signal rather than alphabetically. The grouping matters because a design decision moves a group, not a parameter: choosing a laminate changes Dk, Df and CTE together, and choosing a copper foil changes roughness and conductor loss together.

JobParametersWhere it shows upFixed when?
LossDf, conductor loss, surface roughnessLink budget, range, thermal riseLaminate order and foil choice
Phase and delayDk, TCDk, frequency stability of Dk and DfBeam steering, matched pairs, group delayLaminate and geometry
ImpedanceZ₀ (a result), trace width, dielectric heightReflections, return lossStack drawing, before tooling
ReliabilityCTE, water absorption, thermal conductivityBarrel life, delamination, junction temperatureMaterial family

Two notes before the list. Characteristic impedance is not an independent parameter: it is computed from Dk, height, width and thickness, so treating it as something to be chosen on its own is the most common conceptual error in the field. And water absorption is not the delamination mechanism it is usually said to be on PTFE — the number is 0.02 %, the lowest of any laminate, and the water that blisters a multilayer is held in the bondply and in any FR-4 layers.

1. Dielectric Constant: Two Numbers, Not One

Every high-frequency laminate has a process Dk and a design Dk, and they differ because they are measured differently. The process value comes from a clamped stripline or a similar fixture on the manufactured sheet; the design value is what the supplier recommends for simulation across a frequency band. On RO4003C the two are 3.38 and 3.55. On RO3010 they are 10.2 and 11.20 — a 10 % difference, which is the difference between a working 50 Ω line and an unexplained 47 Ω one.

The sensitivity is what makes Dk worth a section of its own. Impedance scales as the inverse square root of the effective Dk, so a Dk tolerance of ±0.05 on a 3.66 material is about ±0.74 % on impedance; the same absolute tolerance on a 10.2 material is about ±0.35 %, but on a 2.20 material it is ±1.1 %. Tight Dk tolerance costs money and is most worth paying for on low-Dk materials, which is the opposite of the usual intuition.

The other half of Dk is what it buys geometrically. Lower Dk means a wider line for the same impedance, lower conductor loss, and a larger circuit for the same electrical length. Higher Dk means the reverse: 3010 at 10.2 shrinks a resonator to a quarter of the area of a 2.20 material, and pays for it with a narrower line and higher loss per unit length at the same Df-per-electrical-length. Which is why “high Dk is for miniaturisation” and “low Dk is for loss” are both true statements about different problems.

2. Dissipation Factor: The Only Parameter That Scales With Frequency

Df is a ratio, not a loss. The loss it produces grows in proportion to frequency, which is why a material that is acceptable at 1 GHz can be unusable at 10 and hopeless at 77. The working approximation for dielectric attenuation in a homogeneous medium is:

αd ≈ 2.6 · fGHz · √Dk · Df  dB/inch

The formula has one consequence that reshapes most material arguments: at equal electrical length, total dielectric loss is proportional to Df alone. A high-Dk material shortens the trace, and the shorter trace has the same loss as the longer one on low-Dk material with the same Df. High Dk is a size decision; Df is the loss decision.

MaterialDkDf at 10 GHzDielectric loss, dB/in at 10 GHzRelative to 5880
RT/duroid 58802.200.00090.0351×
RO30033.000.00100.0451.3×
RO4003C3.380.00270.133.7×
RO4350B3.480.00370.185.1×
RO301010.20.00220.195.4×
FR-44.20.0201.0831×

Read the 3010 row against the 4350B row: Df is lower, loss is higher, because the trace is narrower and the field is more concentrated for the same impedance. That is not a defect of the material; it is the arithmetic that follows from buying area with Dk.

3. Stability of Dk and Df: Frequency, Temperature and Humidity

A datasheet Dk is a point on a surface. The surface has three axes: frequency, temperature and moisture. A material can be excellent at one point and poor along an axis, and which axis matters depends on the application — broadband systems care about frequency stability, outdoor units about temperature, and anything that is baked or washed about moisture.

MaterialTCDk, ppm/°CDk change per 100 KPhase drift, 100 mm at 10 GHz
RT/duroid 5880−125−1.25 %about 10°
RO4003C+40+0.40 %about 4°
RO4350B / 4835+50+0.50 %about 5°
RT/duroid 6010.2LM−425−4.25 %about 70° — high Dk amplifies it

The phase figures are computed rather than quoted: delay and phase scale with the square root of Dk, so a relative Dk change produces half that change in electrical length. A 100 K excursion at 40 ppm gives 0.40 % on Dk and 0.20 % on phase, which on 100 mm at 10 GHz is a few degrees. The same arithmetic on a 10.2-Dk material gives ten times the Dk shift on a third of the wavelength, which is why a −425 ppm coefficient is a specification item and a ±40 ppm one is a footnote.

The sign matters more than the magnitude in some systems. A negative TCDk means the electrical length shortens as the board heats, so a phase-matched pair built on two different materials can drift in opposite directions — a failure that no room-temperature measurement reveals. For a beamforming array the consequence is a few tenths of a degree of steering error per 100 mm; for a delay-matched filter it can be the whole specification. Whatever the band, the number to demand is the temperature coefficient, not the nominal Dk.

4. Characteristic Impedance: A Result, Not a Parameter

Z₀ is the ratio of voltage to current on a travelling wave, and it is computed, not chosen. What the designer actually controls is three numbers — Dk, dielectric height and line width — and one of them (height) is fixed by the laminate catalogue before the layout starts.

Dk (process)W/h for 50 Ω microstripεeffλg at 10 GHz
2.203.111.8721.9 mm
3.002.542.4219.3 mm
3.382.332.6718.4 mm
3.662.212.8517.7 mm
11.200.857.4111.0 mm

The table carries two practical warnings. First, at 0.254 mm height a 50 Ω line on a 10.2 material is 0.216 mm wide — still manufacturable, but the same ±25 µm of etch error that costs 2.8 % on a 0.56 mm line costs 11 % there. Second, a buried stripline of the same impedance is roughly 2.5 times narrower than the microstrip for the same dielectric height, which is why stripline impedance is more sensitive to every process tolerance: isolation is bought with tolerance sensitivity.

5. Propagation Delay: What Dk Actually Buys You

Delay follows directly from the effective Dk: tpd = √εeff / c. On a 2.20 material with εeff 1.87 the delay is 4.56 ns/m; on FR-4 with εeff around 4.2 it is 6.83 ns/m. The difference over a 100 mm trace is 227 ps, which is small until a matched pair has to hold skew to 10 ps, at which point 45 mm of length mismatch on FR-4 is the entire budget and 68 mm on a 5880 stack.

Two consequences. Low-Dk material buys timing margin as well as loss margin, and on a phased array it buys physical length — a longer line for the same delay, which is often what makes a feed network fit. And because delay depends on εeff rather than on Dk alone, the same laminate produces different delays in microstrip and stripline: the microstrip has air above it, so its effective Dk is lower and its delay per millimetre is shorter. A stack that mixes the two transmission-line types will have two delays per unit length, which is a routing rule rather than a material property.

6. Conductor Loss: Why Widening Stops Helping

Conductor loss falls as the trace widens, but it falls towards a floor, because the dielectric loss underneath does not change. On a 10 GHz microstrip over a 0.0027-Df material the dielectric term is fixed at about 0.13 dB/in while the copper term moves from 0.19 to 0.095 to 0.063 dB/in as the line goes from 0.56 to 1.18 to 1.89 mm.

Line width (50 Ω, matched h)Dielectric, dB/inConductor, dB/inTotal, dB/inGain vs 0.56 mm
0.56 mm (h 0.254)0.130.190.32—
1.18 mm (h 0.508)0.130.0950.22530 %
1.89 mm (h 0.813)0.130.0630.19340 %

The diminishing return is the useful part. Doubling the width once buys 30 % of the total; doubling it again buys another 10 %. Below about 0.5 % of the total, no further widening is worth the board area, and the next lever is the one in the next section.

7. Surface Roughness: 0.65 µm of Skin at 10 GHz

At 10 GHz the current in copper occupies roughly 0.65 µm below the surface. Any foil whose profile is comparable to that depth forces the current along a longer path, and the extra path appears as extra resistance: about +0.1 dB/in on a 0.56 mm line at 10 GHz, rising with frequency and with profile.

Roughness does a second, less-discussed thing: it raises the effective Dk, because the field sees a mixture of copper and dielectric in the roughness layer rather than a clean boundary. That shifts phase as well as loss, which is why a foil change is a re-qualification and not a substitution. The controls are foil specification by profile rather than by weight — HVLP and similar low-profile foils — and a finish that does not add its own roughness. On a 0.0009-Df material, copper is the dominant loss term, so foil choice is the largest single decision after the laminate itself. Foil types are compared in PCB copper foils.

8. CTE: The One You Cannot Match

Coefficient of thermal expansion appears on every list, usually with the advice to match it to copper. In the Z direction that is impossible, and pretending otherwise is how barrel cracks get designed in.

MaterialX / Y, ppm/°CZ, ppm/°CZ expansion, 1.6 mm over 215 K
Copper17176 µm
RO4350B10 / 123211 µm
RO4003C11 / 144616 µm
RO301013 / 11166 µm
Taconic TLY-526 / 1521775 µm
RT/duroid 588031 / 4823782 µm

The mitigation is mechanical rather than material: plating ductility of at least 20 % elongation, an aspect ratio of 6:1 or better, and a symmetric stack — the three controls are worked through in multilayer lamination for high-frequency PCBs. Note also the inversion in the X/Y columns — 4350B and 4003C expand less than copper in plane, which is a registration advantage on large panels; PTFE expands more, which is why panel size and room temperature belong in the dimensional tolerance.

9. Water Absorption: A Small Number With a Real Effect

The absolute numbers are small and the ranking is not what most lists imply. PTFE absorbs about 0.02 % by weight; RO4000 grades about 0.05–0.06 %; FR-4 between 0.1 and 0.2 %. PTFE is therefore the least moisture-sensitive laminate available, and it is not the source of the blistering that gets blamed on it.

What absorption does is shift Dk. Water has a Dk near 80 at low frequency, so even a few hundred parts per million of uptake changes the effective Dk of the composite measurably, which moves phase and impedance. The practical impact depends on the environment and on whether the board is baked before assembly, and the control is storage discipline rather than material selection: sealed bags, first-in-first-out, controlled store, and a bake profile matched to the family that actually absorbs — the prepreg and the FR-4, not the PTFE core.

10. Thermal Conductivity: The Number Everyone Overestimates

Thermal conductivity is the parameter most often used as a proxy for “good thermal design”, and it is the one where the arithmetic is most often skipped. What matters is thermal resistance, which is thickness divided by conductivity — so a low-k material is only a problem in proportion to how thick it is, and a high-k material only helps if the heat actually has to cross it.

Materialk, W/m·KR″ at 0.254 mm, K·cm²/WRise at 10 W/cm²
RT/duroid 58800.2012.7127 K
FR-40.308.585 K
RO4350B0.693.737 K
RO4003C0.713.636 K
RT/duroid 6035HTC1.441.7618 K

The table explains a decision that otherwise looks arbitrary. A 0.254 mm dielectric of 5880 has the same thermal resistance as about 142 mm of copper, so adding a copper layer in the middle of the stack does not shorten the vertical path — the heat still has to cross the dielectric. The solutions that work are the ones that remove the path: thermal vias filled with copper, a metal backing, or a genuinely high-k laminate where the loss budget allows. Treating “more copper” as thermal design is the most expensive misconception in the list. The surrounding argument is in thermal management in PCB manufacturing.

The Ranking: Which Two Decide Most Boards

If the ten are sorted by how much of the outcome they control per unit of effort, the order is not the order of the list. Two parameters decide most boards, three more decide the rest, and the remainder set boundaries rather than outcomes.

RankParameterLeverageWhy
1DfUp to 30× the dielectric lossScales with frequency; nothing else in the list moves that far
2Surface roughness and copper treatmentAbout 0.1 dB/in at 10 GHz, and the dominant term on low-Df materialOnce dielectric loss is small, copper is what is left
3Dielectric height tolerance±6 % impedance for ±10 % heightThe largest single term in the impedance budget
4TCDkDegrees of phase per 100 KDecides arrays and matched pairs, invisible at room temperature
5Z-CTE82 µm on 1.6 mm for PTFEDecides barrel life, not signal performance
6–10Dk, thermal conductivity, water absorption, delay, impedanceSecondaryBoundaries, or consequences of the five above

The practical use of the ranking is allocation. On a 6 GHz link, Df and roughness are worth paying for and TCDk is not. On a 77 GHz radar front end all five matter and the dielectric height tolerance becomes the binding constraint. On a high-Dk miniaturised filter, Dk accuracy and etch tolerance dominate and Df is almost irrelevant, because the line is short.

What Is Missing From the Ten

A list of ten always omits something, and three omissions cost more than two of the entries. Glass weave effect. On glass-reinforced laminates the Dk varies between bundles and windows, at a period of a few hundred micrometres, so two supposedly identical traces can differ in delay by a few picoseconds depending on where they sit relative to the weave. Spread-glass and flat-glass constructions exist precisely for that. The finish as an electrical layer. 3–6 µm of nickel under 0.05 µm of gold puts the current in nickel above a few gigahertz; the finish is part of the transmission line, not a coating on it, as PCB surface treatment processes sets out in detail. The solder mask. Dk 3.5–4.2 and 20–30 µm thick costs 1–3 Ω, and the legend ink at Dk near 6.5 over an RF trace is worse.

All three share a property with the ten: they are chosen at the stack review and expensive to change afterwards. Which is the real argument for treating this as a review checklist rather than a reference list — the numbers are only useful before the drawing is released. The physics underneath them is in concepts of electricity in PCB, and the measurement side in electrical characteristics in high-frequency PCB manufacturing.

Conclusion

The ten critical parameters are four jobs in disguise. Df and copper roughness decide loss; Dk, TCDk and frequency stability decide phase and delay; impedance is computed from Dk, height and width and cannot be specified independently; CTE, water absorption and thermal conductivity decide whether the board survives assembly and service. Once they are grouped that way, a design decision maps to a group and the argument becomes specific.

Ranked by leverage, two parameters dominate: Df, because it is the only term that scales with frequency, and copper roughness, because on a low-Df laminate it becomes the largest remaining loss. Third is dielectric height tolerance, which is the largest term in the impedance budget. The rest set boundaries — and among those boundaries, the one most often misused is thermal conductivity, where a 0.254 mm dielectric of 5880 equals about 142 mm of copper in thermal resistance and no amount of extra copper in the middle of the stack changes the path.

XCE PCB builds 2 to 72 layers across Dk 2.20 to 10.2 on Rogers, Taconic, F4B and FR-4 families, panels to 600×1200 mm, impedance ±10 % with coupon verification, samples in as fast as 24 hours, inspected to IPC-A-600G and IPC-6015. Send the link budget and the frequency band; two of the ten usually turn out to bind, and it is worth knowing which two before the laminate is ordered. The parameter most often argued about on its own is covered in dielectric constants in high-frequency PCBs.

Frequently Asked Questions

Which of the ten parameters matters most?
Dissipation factor, because it is the only parameter whose effect scales in proportion to frequency — a material that is acceptable at 1 GHz can be unusable at 10. Second is copper surface roughness, which becomes the dominant loss term once Df is small: about 0.1 dB/in at 10 GHz on a 0.56 mm line. Third is dielectric height tolerance, worth ±6 % of impedance for ±10 % of height.
Why is characteristic impedance not really a parameter?
Because it is computed from Dk, dielectric height, trace width and copper thickness. The designer controls those four, not Z₀ directly. Specifying 50 Ω without specifying the measurement frequency, the coupon geometry and the layer it applies to leaves the number undefined — and roughly 2.5 times narrower for a buried stripline than for a microstrip at the same height.
How much delay does low-Dk material actually save?
Delay is √εeff divided by c. On a 2.20-Dk laminate with εeff 1.87 that is 4.56 ns/m; on FR-4 with εeff near 4.2 it is 6.83 ns/m. Over 100 mm the difference is about 227 ps, which matters when a matched pair has to hold 10 ps — that is 45 mm of length match on FR-4 against 68 mm on the low-Dk stack.
Does a low thermal conductivity make a laminate unusable for power designs?
Only if the heat has to cross the dielectric. Thermal resistance is thickness divided by conductivity: 0.254 mm of 5880 is 12.7 K·cm²/W, equivalent to about 142 mm of copper, so adding copper in the middle of the stack does not shorten the path. Thermal vias, a metal backing or a high-k laminate like 6035HTC at 1.44 W/m·K are what actually change the number.
Why is PTFE not the moisture-sensitive material it is said to be?
Because it absorbs about 0.02 % by weight, the lowest of any common laminate. The water that blisters a multilayer is held in the bondply, the prepreg and any FR-4 layers, at 0.1–0.2 %. Water does shift Dk, since its own Dk is near 80, but the bake that fixes it belongs to the prepreg, not to the PTFE core.
What is missing from the standard list of ten?
Three things that cost more than two of the entries: glass weave effect, which makes two identical traces differ in delay depending on their position relative to the weave; the surface finish, where 3–6 µm of nickel puts the current in nickel above a few gigahertz; and the solder mask, at Dk 3.5–4.2 and 20–30 µm thick, worth 1–3 Ω on a 50 Ω line.
RF & Microwave PCB Fabrication

Which Two of the Ten Bind Your Design?

Send us your Gerber files and requirements for a manufacturability review and a tailored quotation.

2–72 layers ±10% impedance tolerance Manufacturing since 2009