What Are the Differences Between High-Frequency and Low-Frequency PCBs?

What are the differences between high-frequency and low-frequency printed circuit boards? Most answers list materials, then design rules, then processes, as if the two boards were parallel product lines. They are not. The same FR-4 panel is a low-frequency board or a high-frequency board depending on how long the trace is and what the signal has to survive. What actually changes is that three physical effects switch on at three different points — the return current moves off the shortest resistive path at a few hundred kilohertz, the trace stops behaving like a wire at a tenth of a wavelength, and the dielectric becomes a loss element once the line is long enough. This article locates those three boundaries with numbers, and shows what each one forces you to change.

“Low frequency” and “high frequency” are used to describe three unrelated things: where the return current flows, whether the trace is electrically long, and how much the dielectric attenuates. Each one switches on at a different condition, and only the third has anything to do with the laminate’s loss factor. Lumping them together is the reason so many comparisons produce rules that do not survive contact with a real board.

Take a 100 MHz clock on a 300 mm board. Two hundred kilohertz is nowhere in sight, the wavelength in FR-4 is 1.65 m so the trace is well inside λ/10, and the dielectric loss over that length is a few hundredths of a decibel. This is a low-frequency board in every sense. Now move the same clock onto a 2 m backplane: still 100 MHz, still FR-4, but the line is now electrically long — 1.2 wavelengths — and it needs termination, impedance control and a proper return plane. Nothing about the material changed. Everything about the design did.

That is the correct way to read the two categories: not as a frequency band, but as a set of thresholds that a given trace either crosses or does not.

At DC, current distributes itself to minimise resistance, which means it takes the widest, thickest path available — the whole ground plane, not the line under the trace. As frequency rises, inductance takes over and the current concentrates under the signal trace, because that is the loop with the smallest area. The crossover is where the inductive reactance of the return path equals its resistance, at roughly f ≈ R/(2πL).

For a 1 mm wide, 1 oz trace at 0.49 mΩ per millimetre with about 0.3 nH per millimetre of loop inductance, that works out to roughly 260 kHz. On a plane with lower inductance the number moves up, on a narrow return it moves down, and the practical band is a few hundred kilohertz. Below it, splitting a ground plane is harmless; above it, a split under a signal turns the return into a long detour, and the loop area — and the emission — grows with it.

This is the divide that most often gets missed, because it happens far below any frequency anyone would call high. A 1 MHz switching converter already has its return current following the trace, and a 20 kHz audio circuit does not. The design consequence is the same one that governs 5 GHz boards: never route a signal across a plane split. The physics is a function of frequency, not of the product category.

A trace behaves as a lumped connection while its length stays below about a tenth of the wavelength in the dielectric. Above that, voltage and current vary along the line, reflections appear at any impedance change, and the trace has to be terminated. The wavelength is set by the effective permittivity: on FR-4 microstrip ε_eff ≈ 3.3 gives a velocity of about 1.65 × 10⁸ m/s, on 5880 with ε_eff ≈ 1.87 it is 2.19 × 10⁸ m/s — a third faster, so the same physical length is electrically shorter.

The table explains two things that are otherwise mysterious. First, why a 1 GHz board can be built with no impedance control at all when every RF net is 8 mm long: nothing on it is electrically long. Second, why a 100 MHz backplane needs controlled impedance, termination and a continuous reference plane: at 1.65 m per wavelength, 300 mm is a fifth of a wavelength, and a fifth is enough to reflect.

The practical rule that follows is about geometry. Move a net from 200 mm to 20 mm and it stops needing a transmission-line treatment even if the frequency never changed; move it from 20 mm to 200 mm on a 5 GHz design and it needed that treatment long before the length tripled.

The third divide is the only one that is a property of the material, and it is the one that appears in the invoice. Dielectric attenuation rises with frequency and with the loss factor, and only becomes a budget item once the line is long enough for the product to matter. At 10 GHz, FR-4 spends about 1.08 dB/in in the dielectric while 4350B spends 0.18 and 5880 spends 0.035 — but a 10 mm RF feed on FR-4 costs 0.43 dB and might be entirely acceptable.

This is where the three divides interact. An electrically long line on a lossy laminate is the expensive case: it needs controlled impedance, termination and a low-loss material. An electrically short line on the same laminate needs none of them, however high the frequency. The decision procedure for choosing the laminate from that loss figure is set out separately; here the point is only that loss is the third gate, not the first, and that it is a function of length as much as of frequency.

Placed together, the thresholds show why “high-frequency board” is a description of a design rather than of a product family. Each row has its own trigger, its own symptom and its own fix.

The ordering matters for troubleshooting. Ringing on a 300 mm line is almost never a laminate problem, and swapping to a low-loss material to fix it wastes money without touching the cause. Conversely, an eye that closes steadily with length on a properly terminated 400 mm link is almost never a layout problem.

The original comparison states that high-frequency boards hold ±5 % impedance while low-frequency boards do not need impedance control. The second half is right; the first is optimistic as a default. ±10 % is the standard window, ±5 % requires thickness and etch control on the same stack, and ±3 % requires coupon plus VNA verification per build. Whether the tighter number is worth paying for depends on what a mismatch actually costs at the interface.

Two readings follow. A ±10 % process guarantees no better than 26 dB return loss at the worst case, which is why an antenna port quoted at 20 dB needs the tighter window, not a better connector. And the tolerance budget on FR-4 is already about ±6.6 % before etching, so ±5 % on FR-4 is not a process improvement but a different material and a different press schedule. The stack-up arithmetic behind that figure is in high-frequency stack-up design, and the measurement side in impedance control in manufacturing.

The claim that high-frequency boards use HVLP or reverse-treated foil is correct, but the reason is geometric: the current only occupies a thin layer at the surface, and once that layer is thinner than the copper profile the current has to follow the topography. Skin depth falls as the inverse square root of frequency.

The 0.1 dB/in figure at 10 GHz is worth putting in perspective: it is about the same as the entire dielectric loss of a 4350B line, and about a third of the difference between FR-4 and 4350B. Foil choice is therefore a second-order decision that becomes first-order exactly where the dielectric has already been optimised. On a 2.4 GHz board with 15 mm traces it is unmeasurable; on a 28 GHz antenna feed it is part of the design.

Surface finish belongs to the same argument. At 10 GHz the skin depth in nickel is 1.33 µm, so an ENIG layer 3–5 µm thick carries essentially all the current with about twice the resistivity of copper. That is why OSP or immersion silver is preferred on high-frequency boards and ENIG is preferred where solderability and shelf life matter more.

The usual comparison stops at D_k and D_f. The tolerance on those numbers matters at least as much, because it is what converts a design into a yield. FR-4 is quoted as a range — D_k 4.2–4.6, D_f 0.018–0.025 — while RF laminates are quoted as nominal ± tolerance, which is a different kind of promise.

The one line that surprises designers is TCD_k. Its absence on FR-4 is not an omission in the data sheet; it is the absence of a specification, and a parameter that is not specified cannot be guaranteed, inspected or used in an analysis. Any comparison that treats FR-4 as “the same but lossier” ignores the property that decides phase stability.

Most of the process is identical, and the marketing version of this comparison overstates the difference. Four stations genuinely change.

Layer-to-layer registration. A high-frequency multilayer is normally held to about ±25 µm, because a misregistered RF layer shifts impedance and coupling together. On a conventional digital board the same error shows up as a slightly different breakout, which the router absorbs.

Drilling. PTFE-based laminates are soft, dimensionally unstable and abrasive-filled. They need sharp geometry, reduced stack heights and no laser, since the material decomposes above roughly 400 °C and releases hydrogen fluoride. Thermoset RF laminates like 4350B drill much like FR-4.

Plated through holes. PTFE has no usable adhesion to electroless copper, so hole walls need sodium-naphthalene or plasma treatment first — a step FR-4 never needs. This is the single biggest process difference between the two high-frequency families, and it is invisible in any D_k/D_f table.

Etching and imaging. Chemically identical, but the width tolerance matters more, because the impedance window on a thin-dielectric RF stack is narrower. A 25 µm etch variation is a few percent of impedance on a 0.6 mm line and much more on a 0.24 mm one. See drilling and surface treatment for the station-level detail.

Product categories are a consequence of the three divides, not a definition of them. The table below states, for each common application, which divide it crosses — which is also why the laminate requirement varies so much inside one category.

The last row is the one to keep in mind when a customer asks whether their product is “high frequency”. A backplane at 25 Gb/s crosses two divides and pays for a premium laminate; a 6 GHz Wi-Fi module with 12 mm traces crosses them only in a small area. The label does not predict the BOM. The length and the loss budget do.

Because the label describes a product rather than a set of conditions, it is misapplied in both directions, and both cost money.

Buying RF laminate for an electrically short board. A 2.4 GHz module with a 15 mm antenna feed and a 50 Ω line on FR-4 loses 0.15 dB in the dielectric, and its impedance can be held to ±10 %. Specifying 4350B there buys perhaps 0.1 dB and a tighter D_k, at three to five times the material cost and, on a PTFE part, a different drilling and plating route with a longer lead time. The performance gain is real but smaller than the loss of one badly placed via.

Keeping FR-4 on a long loss-critical net. The opposite error is more expensive. A 400 mm link at 2.5 GHz spends about 4.2 dB in FR-4 dielectric alone; on 4350B it spends 0.68. If the channel budget was written for the second figure and the board was built in FR-4, the failure appears as intermittent link errors at temperature extremes, and it is usually first blamed on connectors, then on the receiver.

There is a cheap way to avoid both: write the three thresholds into the design review checklist as numbers rather than as adjectives — the longest net at the highest frequency, its length against λ/10, and its computed dielectric loss against the budget. A design that answers those three questions has already been classified correctly, and the laminate choice follows from the answer instead of preceding it.

The differences between high-frequency and low-frequency PCBs are not a frequency band. They are three thresholds, and each one is calculated from the design rather than assumed from the product: the return current starts following the trace at a few hundred kilohertz, the trace becomes a transmission line at a tenth of a wavelength — 165 mm on FR-4 at 100 MHz, 16.5 mm at 1 GHz — and the dielectric becomes a budget line once the loss over the net is a meaningful fraction of the allowance.

The material follows from those answers. Short and loosely specified: FR-4, with proper grounding and no impedance control, and no apology. Electrically long: controlled impedance and termination first, laminate second. Long and loss-critical: a low-loss laminate, and a specification that quotes D_k, D_f and TCD_k with tolerances rather than ranges. Phase-critical across temperature: only a material that publishes TCD_k can even be considered.

XCE PCB builds both sides of the line and the hybrid in between: 2 to 72 layers, mixed stacks with sequenced lamination, impedance to ±10 % as standard and tighter on request with coupon and VNA verification, panels to 600×1200 mm, samples in as fast as 24 hours, inspected to IPC-A-600G and IPC-6015. For the loss arithmetic that decides which laminate, see where FR-4 stops working; for the stack that holds the impedance, see high-frequency PCB design steps.