How to Design High Frequency PCB Circuit Boards: Four Decisions That Lock Everything Else

Learning how to design high frequency PCB circuit boards is usually presented as a list of good practices — short traces, smooth corners, low-loss laminate, controlled impedance. In practice the work is decided by four choices made early, and every later choice is a consequence. Pick the band, and the material family follows. Pick the material, and the dielectric height follows from the loss budget. Pick the height, and the trace width is no longer a design variable at all. This article walks those four decisions with the arithmetic at each step, and then corrects six statements about high-frequency design that keep being repeated on the internet and do not survive a data sheet.

RF PCB for High-Frequency Wireless Communication

What the Word Frequency Actually Decides

“High frequency” is not a property of the laminate. It is a statement about how much of the design is decided by electrical length. Below roughly 1 GHz, a board is a set of connections and the laminate is mostly a mechanical carrier. Above roughly 3 GHz the trace is a transmission line whose width, height and copper texture all matter. Above roughly 20 GHz the vertical structures — vias, connectors, the solder mask over the line — become the dominant discontinuity, and the horizontal routing is comparatively easy.

The single most useful table in high-frequency design is therefore a table of wavelengths, because every rule of thumb in the field is a wavelength rule in disguise. A via fence at 0.5 mm spacing is a wall at 10 GHz and a picket fence at 77 GHz; a 3 mm keep-out from the board edge is a quarter wavelength at 25 GHz.

FrequencyFree-space λGuided λ (Dk 3.3 microstrip)First constraint that binds
1 GHz300 mm184 mmAlmost nothing — FR-4 is adequate for short runs
3 GHz100 mm61 mmDf: FR-4 at 0.02 costs 0.32 dB/in of dielectric loss
6 GHz50 mm31 mmDk tolerance and dielectric thickness tolerance
10 GHz30 mm18 mmCopper roughness, then Df
20 GHz15 mm9.2 mmVia stub, connector launch, solder mask
28 GHz10.7 mm6.6 mmTCDk and etch tolerance; λ/20 is 0.33 mm
77 GHz3.9 mm2.4 mmMaterial homogeneity and mechanical registration

Read the right-hand column as a schedule of failures. A design team that spends its effort on routing while the binding constraint is copper roughness will ship a board that misses its loss budget by a factor of two and never find the cause in the layout.

Decision One: The Highest Frequency Is Not the Clock

The first question is not “what is my clock” but “what is the highest significant harmonic”. A digital edge is broadband, and the bandwidth that matters is set by the rise time rather than the repetition rate: BW ≈ 0.35 / tr. A 100 ps edge carries energy to about 3.5 GHz; a 35 ps edge to about 10 GHz; a 20 ps edge to 17.5 GHz. A 1 GHz clock with a 35 ps driver is a 10 GHz board with a slow repetition rate, and must be designed as one.

For RF the equivalent statement is about the modulated bandwidth rather than the carrier, but the carrier still sets the passive behaviour. The practical rule is to design to the highest frequency at which significant energy exists, then apply the wavelength table at that frequency. Designing a 77 GHz radar front end to 10 GHz practices — microstrip on 0.5 mm laminate, through-hole launches, solder mask over the feed — produces a board that works in simulation and fails in the chamber.

One consequence is worth stating early because it determines the rest of the budget: the loss difference between materials is a per-unit-electrical-length quantity, not a per-unit-physical-length one. A low-Df laminate buys you a longer allowable run at the same loss, not a lower loss at the same length — unless you also keep the same physical length, in which case the gain is exactly the ratio of Df values.

Decision Two: The Material Family, Chosen by Loss per Inch

Laminates are sold by brand and sorted by chemistry, but the design decision is a number: dielectric loss per inch at the operating frequency, which follows the standard approximation αd ≈ 2.6 · fGHz · √Dk · Df dB/in. At 10 GHz the whole market collapses into a single column.

Material familyDkDf @10 GHzDielectric lossCost multipleWhere it belongs
FR-44.30.0201.08 dB/in1×Below 3 GHz, short runs, digital
RO4003C (hydrocarbon / ceramic, thermoset)3.380.00270.13 dB/in3–5×3–24 GHz, FR-4 process compatible
RO4350B (same chemistry, V-0)3.480.00370.18 dB/in3–5×Power amps, automotive radar, medical
RO3003 (PTFE, ceramic filled, no glass)3.000.00100.045 dB/in8–15×Low-loss feeds, 24–40 GHz
RT/duroid 5880 (PTFE, micro-glass)2.200.00090.035 dB/in10–20×Millimetre wave, instrumentation
RO3010 (PTFE, high Dk)10.20.00220.18 dB/in10–20×Size-constrained filters and antennas

Two readings matter. First, the jump from FR-4 to RO4003C is a factor of eight in dielectric loss for a factor of three to five in material price, which is why that step is almost always taken at the first sign of trouble. Second, the step from RO4003C to RO3003 is a further factor of three in loss for a further factor of three in price, and it buys nothing unless the copper loss has already been addressed — because on a 0.0027 laminate at 10 GHz, dielectric and copper loss are roughly equal, and on 0.0010 the copper term dominates.

There is also a chemistry trap. The RO4000 series is a hydrocarbon with ceramic filler and a thermoset cure; it is not PTFE, despite being sold as a high-frequency laminate next to PTFE products. The PTFE families are RO3000 and RT/duroid. The practical differences are cold flow (PTFE creeps under pressure, thermoset does not), drilling and desmear chemistry (PTFE needs sodium naphthalenide or plasma, never permanganate), and Z-axis expansion (PTFE 200 ppm/°C and above, thermoset 30–50). Choosing “a PTFE board” when you meant RO4350B changes the entire process route. The pairwise differences are set out in 4350B versus 4003C and in the TLY-5 parameter notes for the PTFE side.

Decision Three: Dielectric Height, Which Sets the Width

Once the laminate is chosen, the dielectric height is the last free geometric variable. On a 50 Ω microstrip the width is not chosen independently; it follows the ratio of width to height, and that ratio is a property of Dk. On Dk 3.38 the ratio is about 2.33, so 0.254 mm of dielectric gives a 0.59 mm line and 0.813 mm gives 1.89 mm.

DkW/h for 50 ΩWidth at h = 0.254 mmWidth at h = 0.508 mmConsequence
2.203.110.79 mm1.58 mmWidest lines; best copper loss, worst density
3.002.540.65 mm1.29 mmThe common microwave compromise
3.382.330.59 mm1.18 mmFR-4 process compatible
3.662.210.56 mm1.12 mmUse design Dk, not process Dk
10.200.850.22 mm0.43 mmFilters shrink; etch tolerance now dominates

The right-hand end of that table is where designs fail quietly. A 0.22 mm line on a high-Dk laminate carries the same impedance as a 0.59 mm line on RO4003C, but ±25 µm of etch error is 11 % of its width against 4 % of the wider line — and 11 % of width is roughly 5 % of impedance. The same fabricator, holding the same etch window, will report compliance on one stack and failure on the other, and be right both times. The geometry itself is covered in microstrip and stripline design; what happens to that width after etching is in route width in PCB manufacturing.

Height also decides the vertical problem. Thick dielectric means wide lines, low copper loss, and long via stubs; thin dielectric means narrow lines, high copper loss (in absolute terms per unit length), and short stubs. At 20 GHz and above, a 1.6 mm through via is a quarter-wave stub at about 23 GHz in FR-4 and 32 GHz in PTFE, which is why the millimetre-wave stack is built thin, with back-drilling, blind vias or via-in-pad rather than with careful routing.

Decision Four: Copper Foil and Surface Finish

Copper is where the design stops being geometry and becomes texture. At 10 GHz the current occupies the outer 0.66 µm of a 35 µm foil, so the surface the current sees is the foil’s tooth profile, not its nominal thickness. Standard electrodeposited foil has a peak-to-valley roughness of 8–10 µm, comparable to fifteen skin depths, and that geometry lengthens the current path and adds 20–45 % to conductor loss. Very-low-profile foil at Rz ≤ 2 µm removes most of that penalty, which is why the copper callout belongs in the same paragraph as the laminate callout rather than in the fabrication notes.

ItemTypical windowWhat it costs at 10 GHz
Standard ED foilRz 8–10 µm+20–45 % on conductor loss, i.e. +0.04–0.08 dB/in on a 0.59 mm line
HVLP / VLP foilRz ≤ 2 µmReference conductor loss; +10–20 % on laminate price
ENIG finishNi 3–6 µm, Au 0.05–0.10 µmCurrent flows in nickel (skin depth 1.33 µm): +0.10–0.15 dB/in
Immersion silver0.12–0.40 µmNo nickel barrier: lowest loss of the solderable finishes
OSP0.2–0.5 µmOrganic, no metal added; best for wire bonds after cleaning only
Solder mask over RF line20–30 µm, Dk 3.5–4.2Lowers impedance by 1–3 Ω and adds a lossy layer

The finish row is the one most often left to the purchasing department. A 0.05–0.10 µm gold layer is eight to fifteen times thinner than the skin depth in gold at 10 GHz, so it does not conduct; the 3–6 µm of nickel beneath it is two to five times thicker than the skin depth in nickel, so the current is entirely in a metal with 2.04 times the resistivity of copper. On a 25 mm feed that is a few tenths of a decibel — measurable, tolerable. On a 150 mm corporate feed network it is the difference between meeting and missing the budget. The finish windows are tabulated in surface treatment on high-frequency PCB, and the copper side in copper thickness in PCB.

What Each Band Does to the Stack

The four decisions resolve into a small set of stack archetypes. The table is deliberately coarse, because the actual layer count depends on routing, not on frequency; what frequency changes is which layers can be trusted with the sensitive nets.

Operating bandPractical stackWhyNon-negotiable detail
Up to 3 GHz4 layers, FR-4, microstrip on outerLoss budget rarely binds; impedance tolerance doesContinuous ground under every RF line; no split under a crossing
3–10 GHz4–6 layers, RO4003C / RO4350BDf now costs more than the material premiumSolder-mask pull-back on the RF nets; HVLP copper
10–28 GHz6–8 layers, RO3003 or RO4003C, thin coreStub resonance and roughness overtake DfBack-drilled or blind vias; λ/20 stitching at 0.33 mm
28–77 GHz8 layers, PTFE 5880 / 3003, 0.127–0.254 mm coreEverything is a discontinuity; tolerance is the productVia-in-pad, no mask on RF, ±25 µm registration, connector launch designed as a structure

Notice what is absent: there is no archetype in which a metal-core substrate appears. Metal-backed boards exist to move tens of watts out of LEDs and power modules; a 77 GHz radar front end dissipates a few watts at most, and its problem is phase stability, not junction temperature. Putting a radar laminate on aluminium adds a CTE mismatch of roughly 23 against 10 ppm/°C and buys nothing the copper plane was not already doing.

Six Statements That Do Not Survive the Data Sheet

The original version of this article is a useful catalogue of the genre’s recurring errors, because it contains six of them in four hundred words.

StatementWhat the data sheets say
PTFE materials such as the Rogers RO4000 seriesRO4000 is a hydrocarbon with ceramic filler and a thermoset cure. PTFE families are RO3000 and RT/duroid. The process routes differ at desmear, drilling and lamination temperature
RO4000 Df as low as 0.001RO4350B is 0.0037 and RO4003C is 0.0027 at 10 GHz. 0.0010 belongs to RO3003; 0.0009 to RT/duroid 5880
Line width tolerance ±5 µm, registration under 50 µm±5 µm is a laboratory figure. Production on 1 oz copper is ±13 µm (0.5 mil) on finished width, and registration is ±25 µm per lamination cycle — so a three-cycle build accumulates ±75 µm
Above 100 °C choose a laminate with Tg ≥ 180 °CPTFE has no glass transition in the useful range (Tg is about −120 °C) and is used to 250 °C. The criteria that matter are Td, the melting point for PTFE (about 327 °C) and Z-axis expansion
77 GHz automotive radar relies on a metal substrate for heatMetal-core boards serve LED and power stages. Radar modules dissipate single-digit watts; their constraint is phase stability and TCDk, which is why they use PTFE or hydrocarbon laminates
Multi-point grounding for digital, single-point for RFSingle-point grounding is only valid below about 1 MHz, because above that the interconnect inductance dominates. At RF the reference plane is continuous and the return current chooses its own path directly beneath the trace

The pattern behind all six is the same: a rule that was true in one band, or for one chemistry, is repeated as a general law. High-frequency design is unusually prone to this because the same words — loss, ground, tolerance — mean different things at 100 MHz and 30 GHz.

The Numbers a Fabricator Needs Before Quoting

A design package that carries these nine items can be quoted and built without a clarification loop; a package missing three of them will be quoted either expensively or wrongly.

1. Highest significant frequency, stated as a number, along with the rise time if the board is digital. 2. Impedance targets per line type — single-ended, differential, coplanar — with the tolerance you actually need, not the tolerance you would like. 3. Laminate and copper foil per layer, with the foil roughness class named. 4. Dielectric height per layer, as a target with a window, and the layer pair it applies to. 5. Finished copper weight per layer, distinguishing base foil from plated outer copper. 6. Finish and its thickness window, plus where mask must be pulled back. 7. Via strategy — through, blind, back-drilled, filled — with the stub length you will accept. 8. Coupon definition, including coupon layer, length and reference plane. 9. Acceptance method: coupon TDR, VNA, cross-section, or all three, and the sample rate.

Item seven is the one that quietly changes the price. Back-drilling adds a drilling and inspection step per via; filled and capped vias add plating, planarisation and X-ray; blind vias add lamination cycles, and every extra cycle adds ±25 µm of registration error on top of the previous ones. A stack that needs two extra cycles to save 0.05 dB/in is usually a bad trade at 10 GHz and a good one at 40 GHz. How the fabricator holds items two, four and nine is what impedance control in manufacturing is about.

Where the Board Edge and the Cavity Fit In

Edge treatment is the part of high-frequency design that survives longest in folklore and shortest in measurement. The claim that signals must be kept 3 mm from the board edge is not wrong so much as arbitrary: what matters is whether the field of the trace is contained by the reference plane. A microstrip 0.254 mm above its ground plane has most of its field within about 3h — 0.76 mm — of the trace, so a 1 mm margin is adequate at 10 GHz and a 0.2 mm margin is marginal at 77 GHz, where the guided wavelength is 2.4 mm and the edge behaves as a radiating aperture.

Two controls actually change behaviour. Ground stitching along the edge at λ/20 intervals turns the edge into a via fence and suppresses the parallel-plate mode between the top and bottom planes, which is the mechanism behind most edge radiation on thick boards. And where the board sits inside a metal enclosure, the cavity resonance — roughly 150 / L GHz for a dimension L in millimetres — must be kept away from any harmonic of the clock; a 20 mm wide cavity resonates at 7.5 GHz, which lands squarely in the band of a 2.5 GHz clock’s third harmonic.

Simulation: What to Model and What to Skip

Field solvers are cheap and time is not, so the question is where they earn their keep. Three things are worth simulating: the connector-to-trace launch, any filter or matching network whose response must be within a few tenths of a decibel, and the stack as a whole when an impedance target is tight enough that etch tolerance alone would consume the window. Those are the places where an analytic estimate is not accurate enough to decide.

Two things are usually not worth a full-wave model. The first is a straight 50 Ω line of known geometry — the closed-form width is accurate to a few percent and a coupon will tell you the truth. The second is a broad sweep of routing topologies: run one, measure it, and use the result to calibrate a rule, rather than simulating twenty variants of a structure the fabricator will build with different etch compensation anyway.

For digital boards, the analogue of this is eye-diagram simulation with a measured instead of a nominal channel model. A 10 Gbps link simulated with nominal Dk and nominal roughness will show margin that the hardware does not have, because the two errors — Dk low, roughness low — both flatter the result.

Conclusion

How to design high frequency PCB circuit boards, stated as a sequence: fix the highest significant frequency and its guided wavelength; choose the laminate by loss per inch at that frequency, remembering that RO4000 is hydrocarbon and RO3000 / RT/duroid are PTFE; choose the dielectric height, which then fixes the trace width through the W/h ratio; and choose foil roughness and surface finish as electrical decisions rather than purchasing ones. Everything else — via strategy, stitching pitch, mask pull-back, coupon design — follows from those four with arithmetic rather than opinion.

The failure mode to watch is not a bad decision but a decision taken in the wrong order: choosing a laminate before knowing the band, or drawing a width before choosing the height, produces a board that cannot be fixed in layout. The other failure mode is inherited folklore — single-point grounding, PTFE for everything above a gigahertz, metal cores for radar — which is harder to remove because it looks like experience.

XCE PCB builds to this arithmetic on both process routes: hydrocarbon and PTFE, 2 to 72 layers, 0.127 mm cores and up, impedance held to ±10 % with coupon TDR on every panel, panels to 600×1200 mm, samples in as fast as 24 hours, inspected to IPC-A-600G and IPC-6015. The design sequence itself is set out in high-frequency PCB design steps, and the stack arithmetic in high-frequency PCB stack-up design.

Frequently Asked Questions

Is Rogers RO4000 series PTFE?
No. RO4000 materials such as RO4003C and RO4350B are hydrocarbon laminates with ceramic filler and a thermoset cure. The PTFE families are RO3000 (ceramic-filled, no glass) and RT/duroid, including 5880. The distinction matters in fabrication: PTFE creeps under lamination pressure, needs sodium naphthalenide or plasma for desmear and plated-through holes, and expands 200 ppm/°C and more in Z, while RO4000 processes much like FR-4.
What frequency makes FR-4 unusable?
There is no single threshold, because what matters is loss per unit length against the budget. FR-4 at Df 0.02 loses about 1.08 dB/in of dielectric loss at 10 GHz, so a 100 mm run costs roughly 4.3 dB before copper loss, which is usually unacceptable. At 3 GHz the same material costs about 0.32 dB/in, which is fine for short runs. The crossover for most designs is between 3 and 6 GHz.
How wide should a 50 ohm trace be?
It follows the ratio of width to dielectric height for the laminate Dk: about 2.33 for Dk 3.38, 2.54 for Dk 3.00 and 3.11 for Dk 2.20. So 0.254 mm of RO4003C gives a 0.59 mm line, and 0.254 mm of RT/duroid 5880 gives 0.79 mm. Width is not chosen independently; if you want a different width you must change the dielectric height.
What line width tolerance is realistic in production?
About ±13 µm (0.5 mil) on finished width for 1 oz copper under normal control, and around ±25 µm for heavy copper or wide panels. Layer-to-layer registration is about ±25 µm per lamination cycle, so a build with three cycles accumulates error rather than repeating it. A ±5 µm callout will be treated as a laboratory requirement and priced accordingly.
Does a metal-core substrate help a 77 GHz radar board?
No. Metal-backed laminates exist to move tens of watts from LEDs and power stages. A radar front end dissipates a few watts, and its limiting factor is phase stability, which depends on TCDk and dielectric thickness tolerance rather than on heat spreading. Aluminium also expands at about 23 ppm/°C against 10–17 for the laminate, adding a CTE mismatch the copper plane was already handling.
Should analog and digital grounds be joined at one point?
Not above about 1 MHz. Single-point grounding is valid only while the connection impedance is dominated by resistance; at RF the inductance of the link dominates and the join becomes a radiator. The working practice is a continuous reference plane with the return current flowing directly beneath its trace, and a partition of function by placement rather than by cutting the plane.
RF & Microwave PCB Fabrication

Send the Four Decisions, Get a Stack and a Quote Back

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

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