Almost every article on this subject answers with a number: high-frequency boards work above 1 GHz, regular boards below. That line does not hold. A 12 mm Bluetooth trace at 2.4 GHz on ordinary FR-4 loses about 0.12 dB, while a 400 mm backplane running a 5 Gb/s signal on the same material loses over 4 dB in the dielectric alone. The two boards are not separated by frequency but by whether dielectric loss, impedance tolerance and phase drift have become line items in the budget. This article replaces the definition with three tests you can calculate, and shows where each one flips the decision.
The conventional answer is that high-frequency PCBs operate at 1 GHz and above, while regular boards run from tens of megahertz to a few hundred. This survives because it is easy to write and easy to test, not because it separates the two families. Modern FR-4 carries 10 Gb/s serial links whose fundamental sits at 5 GHz, and a 100 MHz signal driven down three metres of cable suffers more attenuation than a 10 GHz signal crossing 20 mm of the same laminate.
What actually separates the two is whether the dielectric is a loss element in the link budget, whether its permittivity is specified tightly enough for the impedance target, and whether its temperature behaviour shows up in phase or delay. Those three questions have numbers attached, and the numbers are computed from frequency and length together — never from frequency alone.
The clearest way to see it is to express dielectric loss as attenuation per unit length, which for a homogeneous line is approximately αd ≈ 2.6 · fGHz · √Dk · Df dB/in. Frequency enters linearly; the loss factor enters linearly; the square root of Dk enters as a weak correction. That single expression explains most of what the rest of this article quantifies.
“Regular PCB” in practice means FR-4: woven glass in an epoxy matrix, with a glass transition somewhere between 130 °C and 180 °C depending on grade. “High-frequency PCB” is not one material either. It is at least two distinct chemistries that happen to share low loss: hydrocarbon/ceramic thermosets that run on an FR-4 process, and PTFE with ceramic filler or microglass, which needs its own press and drilling regime.
| Laminate | Construction | Dk (process) | Df at 10 GHz | TCDk ppm/°C | Z-CTE ppm/°C | k, W/m·K | UL 94 |
|---|---|---|---|---|---|---|---|
| FR-4 (Tg 130–180) | Epoxy + woven glass | 4.2–4.6 | 0.018–0.025 | Not published | 50–70 below Tg, 200–300 above | 0.25–0.40 | V-0 typically |
| RO4350B | Hydrocarbon / ceramic thermoset | 3.48 ±0.05 | 0.0037 | +50 | 32 | 0.69 | V-0 |
| RO4003C | Hydrocarbon / ceramic thermoset | 3.38 ±0.05 | 0.0027 | +40 | 46 | 0.71 | None assigned |
| RO3003 | Ceramic-filled PTFE, no glass | 3.00 ±0.04 | 0.0010 | Low | 17 / 16 / 25 | 0.50 | V-0 |
| RT/duroid 5880 | PTFE + microglass | 2.20 ±0.02 | 0.0009 | −125 | 31 / 48 / 237 | 0.20 | V-0 |
Two columns deserve attention because they are usually left out of comparisons. The first is the temperature coefficient of Dk. FR-4 has no published figure at all, which is not an oversight: it is not a specified parameter, so no supplier can be held to one. Every RF laminate quotes it, and the sign matters — +40 ppm/°C on 4003C and −125 on 5880 drift in opposite directions, and the second is the larger number.
The second is the Z-axis coefficient of thermal expansion. FR-4 expands 50–70 ppm/°C below Tg and two to five times that above it, which is why the glass transition is a reliability parameter rather than a marketing one. The 4000-series thermosets sit at 32–46, PTFE sits at 217–237 — twelve to fourteen times copper — and that single number governs plating ductility, aspect ratio and reflow count on PTFE builds.
One material that does not belong in this table is polyimide. It is a superb film for flex circuits and high-temperature assemblies, but its loss factor sits between 0.008 and 0.02, which is no better than FR-4 and often worse. Listing PI among low-loss high-frequency substrates, as the earlier version of this article did, sends designers down a dead end.
With the attenuation expression above, the comparison stops being qualitative. The table below computes αd for each laminate at three frequencies, using the Dk and Df from the manufacturer data sheets, and expresses each as a multiple of FR-4.
| Laminate | Dk / Df | 1 GHz | 10 GHz | 28 GHz | Relative to FR-4 |
|---|---|---|---|---|---|
| FR-4 | 4.3 / 0.020 | 0.11 dB/in | 1.08 dB/in | 3.02 dB/in | 1× |
| RO4350B | 3.48 / 0.0037 | 0.018 | 0.18 | 0.50 | 6× lower |
| RO4003C | 3.38 / 0.0027 | 0.013 | 0.13 | 0.36 | 8× lower |
| RO3003 | 3.00 / 0.0010 | 0.0045 | 0.045 | 0.126 | 24× lower |
| RT/duroid 5880 | 2.20 / 0.0009 | 0.0035 | 0.035 | 0.097 | 31× lower |
Read the last column as a budget, not as a ranking. A factor of six sounds like a lot until it is placed against the length of the net, and that is where the decision really happens. Turned around, the table answers a more useful question: how far can each laminate carry a signal before it has spent 3 dB, the sort of allowance a short microwave feed or a low-power receive path might have.
| Laminate | Distance per 3 dB at 2.4 GHz | At 10 GHz | At 28 GHz |
|---|---|---|---|
| FR-4 | 294 mm | 71 mm | 25 mm |
| RO4350B | 1.77 m | 424 mm | 152 mm |
| RO4003C | 2.4 m | 592 mm | 212 mm |
| RT/duroid 5880 | 9.2 m | 2.18 m | 788 mm |
This is why the same FR-4 can be perfectly adequate in one 2.4 GHz product and hopeless in another. At that frequency it spends 3 dB every 294 mm of dielectric. A module with a 15 mm antenna feed is using 5 % of a modest budget; a distributed antenna system with a 600 mm harness has already exceeded it before any connector is counted.
Rather than compare materials, run the design through three gates in order. Each gate has a threshold and a consequence, and the first one that fails determines the laminate — there is no need to weigh all seven properties at once.
| Gate | What to compute | Threshold | If it fails |
|---|---|---|---|
| 1. Loss | Take the longest net at the highest frequency, multiply αd by its length, add copper loss, and compare with the link budget | Dielectric term above about 20–30 % of the budget | Move to a low-Df laminate, or shorten the net. At 2.5 GHz check the copper share first — on a low-loss laminate it is already over 70 % of the total, so changing Df again buys almost nothing |
| 2. Impedance | Combine Dk tolerance (Z scales roughly as √Dk), dielectric thickness tolerance and etch tolerance | Required window tighter than ±10 % | FR-4 cannot hold it: Dk ±0.2 is ±2.4 % of impedance and thickness ±10 % is ±6 %, which already totals around ±6.6 % before etch. Use a laminate specified to ±0.05 in Dk |
| 3. Phase and temperature | Multiply TCDk by the temperature excursion, convert to a Dk shift, then to electrical length | Phase drift that eats a meaningful part of the beam-steering or timing budget | FR-4 has no published TCDk, so it cannot be assessed at all. 4003C (+40 ppm/°C) and 5880 (−125) drift in opposite directions and by different amounts |
| 4. Thermal | Power dissipated per unit area against the through-thickness thermal resistance t/k of the dielectric | Rise that forces a design change | Note the direction: PTFE is the worst conductor here (0.20 W/m·K against 0.69 for 4350B), so a low-loss choice can create a thermal problem that FR-4 did not have |
| 5. Compliance | Fire rating, halogen content, outgassing and any application-specific clause | A rating is mandatory for the end product | Watch the trap: 4350B carries UL 94 V-0, 4003C has no rating assigned. Electrically the two are close, so this can decide the part number on its own |
| 6. Cost | Laminate multiple × the fraction of board area that actually needs it, plus added press steps | The multiple exceeds the value of the performance gained | Consider a hybrid rather than a full RF stack |
The order matters. Loss decides whether a low-loss material is needed at all; impedance decides whether FR-4 can hold the tolerance even when loss is fine; phase decides between the low-loss candidates; thermal and compliance eliminate options that looked good electrically. Run in the other order — starting from a favourite material and justifying it — the same data produces the opposite conclusion, which is roughly how most of these decisions get made.
Loss is the divide everyone expects. Tolerance is the one that surprises people, because it fails quietly: the board works, and a fraction of units miss the return-loss limit. On a microstrip the impedance depends on width relative to dielectric thickness, so both dimensions carry tolerances, and the permittivity carries a third.
| Contribution | Typical FR-4 | Effect on Z0 | Typical RF laminate | Effect on Z0 |
|---|---|---|---|---|
| Dk tolerance | ±0.2 (4.3 nominal) | about ±2.4 % | ±0.05 (3.38 nominal) | about ±0.74 % |
| Dielectric thickness | ±10 % | about ±6 % | ±10 % (pressed) | about ±6 % |
| Etch, 0.254 mm stack | ±25 µm | ±1–3 % depending on line width | ±25 µm | same |
| Combined, root sum square | — | ±6.6 % before etch | — | ±6.1 % before etch |
The arithmetic shows why ±10 % is the practical floor for FR-4 and why ±5 % on FR-4 is a claim that rarely survives a coupon measurement: thickness tolerance alone consumes 6 % of the window. It also shows that switching laminate does not fix everything — 6.1 % is not much better than 6.6 %. The remaining margin comes from thickness control, which is a press question, not a material question. That is why the same fabricator holds ±5 % on one stack and ±10 % on another without doing anything differently.
There is a second, less obvious consequence. On a controlled-impedance board the line width is not a free variable: at 50 Ω it is fixed by the dielectric thickness at roughly W/h ≈ 2 for FR-4 and 2.2–3.1 for the RF laminates. A design that changes material while keeping the same stack inherits a different width, and a design that keeps the width shifts its impedance. This is the mechanism behind most “we swapped the laminate and the yield dropped” reports, and it is covered in detail in the 4350B versus 4003C comparison and in stack-up design.
For a phased array, a delay line or a filter that must hold its centre frequency over temperature, the relevant material property is not Df but TCDk, the rate at which permittivity moves with temperature. Every RF laminate quotes it. FR-4 does not, and that is the deciding fact: a parameter that is not specified cannot be guaranteed, measured at goods-in, or used in a tolerance analysis.
The magnitude is not small. On 100 mm of 50 Ω line at 10 GHz, the accumulated electrical length is about 2,300° on 4350B and 1,780° on 5880, because the two materials have different Dk and therefore different wavelengths. Across a 200 °C excursion the phase moves roughly 11° on 4350B and roughly 22° on 5880 — the PTFE part is the more temperature-sensitive of the two despite having the lowest loss, because its TCDk is the largest in absolute value. Choosing on Df alone would pick the wrong one for a phase-critical design.
That inversion is worth remembering: on this site’s materials, loss and phase stability do not rank the same way. 3003 at 0.0010 and 5880 at 0.0009 are nearly identical in loss, but 3003 has a small positive TCDk and 5880 has −125 ppm/°C. For a narrowband filter in a temperature-controlled enclosure, either works; for an outdoor array, the difference is the calibration table.
The early version of this article repeated four statements that are common on manufacturer blogs and wrong in ways that change decisions.
| Claim | What is actually true |
|---|---|
| Polyimide is a low-loss high-frequency material | PI has Df of roughly 0.008–0.02, equal to or worse than FR-4. It is chosen for flex, thinness and solder resistance, not for loss. Using it as an RF substrate adds loss and absorbs moisture |
| Microstrip or stripline “reduces EMI” | They are controlled-impedance geometries, not shields. Emissions depend on loop area and common-mode current. A microstrip over a split plane radiates more than a well-referenced stripline, and a badly routed stripline is no quieter than a microstrip. What the geometry changes is field confinement and the impedance you can hold |
| High-frequency boards have lower propagation delay | Delay is set by effective permittivity, not by the frequency rating: tpd ≈ 3.336 · √εeff ns/m. On microstrip that is about 6.06 ns/m on FR-4, 5.45 on 4003C and 4.56 on 5880 — a 25 % reduction, which comes from Dk being lower, not from the board being called high-frequency |
| Regular PCBs stop at a few hundred megahertz | FR-4 routinely carries multi-gigabit SerDes. Those links work because the traces are short and the receiver has equalisation, not because the dielectric became better. The same laminate fails the same loss test at 400 mm |
A fifth claim, that a higher Dk causes higher loss, confuses two different quantities. At fixed electrical length, loss is set by Df alone; a high-Dk material shortens the physical line for the same phase, so its total dielectric loss is unchanged or lower. RO3010 at Dk 10.2 and Df 0.0022 actually loses less per wavelength than 4350B at 0.0037. High Dk is a decision about size, not about loss.
Most designs do not need the whole panel to be low-loss. A 6-layer control board with one 24 GHz front end has perhaps 15 % of its area carrying RF, and cladding the other 85 % in PTFE pays for performance nothing uses. A hybrid stack puts one or two RF cores where the signal lives and FR-4 everywhere else.
The economics are straightforward, the process is not. Three press systems do not co-cure: FR-4 cures at 170–190 °C, the 4000-series thermosets at 190–200 °C, and PTFE with FEP bonding film at 260–300 °C. A hybrid therefore needs sequenced lamination or a bonding layer that bridges the two, and each added press cycle adds registration error — roughly ±25 µm per cycle, which comes straight out of the same tolerance the RF section needs.
Two failures recur often enough to be designed out in advance. The first is an unbudgeted bonding layer: a 0.10 mm bondply of Dk ≈ 3.5 under a 0.254 mm core moves a 50 Ω line from 0.56 mm to about 61 Ω if the stack is written without it — a 22 % error that no amount of impedance tuning at the fabricator can recover. The second is PTFE cold flow: pressing PTFE 4 % thinner takes 50 Ω to 48.8 Ω, and unlike a thermoset the deformation does not come back. Both are stack-up documentation problems before they are manufacturing problems, which is why the design-file package for hybrid boards lists the bondply and the press sequence as deliverables in their own right.
The differences that reach production are narrower than the marketing suggests. Most stations run identically on both families; four do not.
Lamination. Temperature window and pressure profile differ, and PTFE needs a longer, slower cycle to avoid cold flow. Drilling. PTFE is soft and abrasive-filled; it needs sharp geometry, lower stack heights and no laser, because the material decomposes above roughly 400 °C and releases hydrogen fluoride. Plated through holes. PTFE has no usable adhesion to electroless copper, so hole walls need sodium-naphthalene or plasma treatment before plating — a step FR-4 does not require. Etching and imaging. Chemically the same, but the width tolerance matters more because the impedance window is tighter.
Everything else — inner-layer imaging, AOI, solder mask, surface finish, electrical test — is the same equipment with different parameters. The practical consequence is that a fabricator who already runs FR-4 can build a 4350B board on the same line the same week, while a 5880 board needs the PTFE cell. That distinction, not the Df number, is usually what the lead time and the price multiple are telling you. Drilling and lamination are where the differences concentrate.
The tests are easier to trust with numbers attached. The three below use the same attenuation expression and take the link budget as given.
| Case | Longest net at fmax | Budget | FR-4 | Result |
|---|---|---|---|---|
| BLE module, 2.4 GHz | 12 mm antenna feed | 1 dB | 0.12 dB | FR-4 is correct. With copper the total is under 0.2 dB; a low-loss laminate would add cost for 0.1 dB |
| 24 GHz radar front end | 40 mm feed | 2 dB | 4.1 dB | FR-4 fails by a factor of two. 4350B gives 0.68 dB, 3003 gives 0.17 dB; choose on phase stability, not on the last 0.5 dB |
| 5 Gb/s backplane | 400 mm, 2.5 GHz fundamental | 10–12 dB channel | 4.2 dB dielectric | FR-4 is marginal once copper and connectors are added. 4350B gives 0.68 dB. The label says digital, the arithmetic is identical |
The third case is the one worth keeping. The design is not an RF product, nobody would call its board high-frequency, and it is decided by exactly the same calculation as the radar feed. Conversely the first case is unambiguously RF and unambiguously fine on FR-4. Any definition based on a frequency number gets both of these wrong.
Six items turn the comparison into a specification the fabricator can hold you to:
1. Dk with tolerance and test method — process value, not design value, and the frequency at which it was measured. 2. Df at the operating frequency, since it rises with frequency on every laminate. 3. TCDk where phase matters, with a stated temperature range. 4. Stack-up with bondply thickness called out, so the 50 Ω line is computed on the real dielectric height. 5. Impedance target per layer with coupon geometry. 6. Fire rating and any application clause — the difference between a part with UL 94 V-0 and one with no rating is not visible on a data sheet comparison table unless you look for it.
With those six, the material question becomes a check against a coupon and a certificate. Without them, it becomes a debate between two data sheets measured under different conditions.
The differences between high-frequency PCB and regular PCB are not a frequency band. They are three numbers, and they are computed from the design rather than looked up. Loss per unit length, multiplied by the actual net length, tells you whether the dielectric is a budget item: at 10 GHz FR-4 spends 3 dB in 71 mm while 4350B spends the same in 424 mm. Permittivity tolerance tells you whether the impedance target is reachable: FR-4 at ±0.2 in Dk plus ±10 % in thickness uses about ±6.6 % before etch, which leaves no room for a ±5 % spec. And TCDk tells you whether phase will hold, a question FR-4 cannot answer because it publishes no figure.
The material follows from those answers. Short, low-power, loosely specified — FR-4, and there is no shame in it. Long, high-frequency, tightly specified — a low-Df laminate, chosen between thermoset and PTFE on temperature behaviour and process rather than on the loss number alone. Mixed — one RF core inside an FR-4 stack, with the bondply and the press sequence written into the documentation.
XCE PCB builds all three: 2 to 72 layers, mixed dielectric stacks with sequenced lamination, impedance held to ±10 %, panels to 600×1200 mm, samples in as fast as 24 hours, inspected to IPC-A-600G and IPC-6015. Coupons are measured at the panel centre and the edge, and the material certificate travels with the shipment. For the argument that the error budget simply changes owner between the two families, see high-frequency PCB versus standard PCB; for the loss figures behind the copper side of the budget, see copper thickness in high-frequency PCBs and surface treatment processes.
Send us your Gerber files and requirements for a manufacturability review and a tailored quotation.