Six Tips for High-Frequency PCB Routing: What Each One Fixes, and the Number That Decides It

The six tips for high-frequency PCB routing that circulate in every design guide are correct in their conclusions and unreliable in their reasons — which is why they are applied where they do not matter and skipped where they do. Shorten the traces, avoid 90° corners, split the grounds, close the loops, cut the vias, decouple at the pin: six instructions, three of which are still true above a gigahertz and three of which need restating. This article takes each tip in turn, names the mechanism, gives the number that decides whether it matters, and says where the tip stops being true.

Why High Frequency PCB Prototyping Is Essential Before Mass Production

Why the Format Matters More Than the List

A rule of thumb without a frequency range is not a rule; it is a habit. Every tip below is a real effect, but each one becomes significant at a particular electrical size, and below that size the same advice buys nothing while consuming layout effort, layer count or component cost. The audit format is therefore: what the tip is actually fixing, the number that tells you whether the effect is large in your band, and the conditions under which the tip is wrong.

TipMechanismDeciding numberVerdict above 10 GHz
1. Keep traces shortLoss and phase accumulate with lengthdB per 100 mm at the operating frequencyStill true; now the dominant budget item
2. Avoid 90° cornersLocal capacitance at the bendEquivalent extra length, in micrometresMarginal below 20 GHz; matters as line-width continuity, not as angle
3. Split analog and digital groundIsolation of noisy return currentsReturn-current spread, roughly 3× dielectric heightWrong: splitting breaks the return path
4. Avoid large loopsLoop area is the radiating apertureArea ratio, in dBStill true; the primary emission mechanism
5. Reduce layer changes and viasVia inductance and stub resonancenH per via, and the quarter-wave stub frequencyTrue, but the reason is inductance, not the 0.5 pF of capacitance
6. Decouple at the pinMounting inductance sets the useful bandnH per millimetre of connectionTrue, and stricter than the original wording implies

Two of the six are frequency-independent in form and matter more as frequency rises; two are frequently misapplied; and one — the ground split — is actively harmful on a high-frequency board. That is a useful summary of the genre: the advice is old enough that part of it predates the frequencies it is now applied to.

Tip 1: Short Traces, Measured in Wavelengths

“Keep high-frequency traces as short as possible” is true for three reasons that are often conflated: attenuation, phase, and the size of the structure as an antenna. The first is easy to quantify. On a 0.508 mm RO4003C microstrip, a 50 Ω line loses about 0.48 dB per 100 mm at 2.5 GHz, 1.24 dB at 10 GHz and 2.72 dB at 28 GHz. Halving the length halves the loss, and that is the entire argument for layout compaction — no mystery, just a per-length price.

The second reason is phase. At 10 GHz the guided wavelength on that stack is about 18 mm, so 100 mm of trace is 5.4 wavelengths. A length change of 1 mm is 20° of phase, which matters for any feed network, phase shifter or calibrated cable. This is where “short” becomes a different instruction: not “as short as convenient” but “equal to a tolerance”. On a four-element array at 10 GHz, matching the four feeds to within 1 mm keeps the phase error inside 20°; matching them to 0.2 mm keeps it inside 4°.

The third reason is that a trace becomes an antenna when it approaches a quarter wavelength. At 28 GHz, a quarter wavelength in the dielectric is about 1.6 mm — a stub of that length is a radiator. Below 3 GHz the same statement requires 15 mm, which is why “short” feels like a soft constraint on low-band boards and is a hard one on millimetre-wave boards.

Tip 2: Corners, the Weakest of the Six

The advice to replace 90° corners with 45° chamfers or arcs is the most repeated and the least consequential of the six. A right-angle bend in a 50 Ω microstrip adds a small amount of capacitance — equivalent to a fraction of a millimetre of extra line — and removes a small amount of length. Below about 20 GHz the resulting impedance discontinuity is smaller than the discontinuity from the etch tolerance of the line itself, and it is difficult to measure on a coupon.

What actually causes the reflection that gets blamed on the corner is a width change at the bend. A trace that enters a corner at 0.59 mm and leaves at 0.52 mm, because the CAM tool applied a different compensation to the diagonal segment, has a 12 % width step and a real impedance discontinuity. Mitring the corner does not fix that; keeping the width constant does. The practical rule is therefore: use 45° or arcs because they are easier to keep at constant width and easier to route without neck-down, not because the angle itself radiates. Why finished width is not the drawn width is set out in route width in PCB manufacturing.

Two cases do justify deliberate corner geometry. Above about 40 GHz, the corner’s own capacitance becomes a measurable fraction of the line’s total, and a properly mitred or swept bend recovers a few tenths of a decibel. And in differential pairs, a corner routed with the two lines at constant gap requires an arc; a 90° corner with a constant gap is geometrically impossible and produces an intra-pair skew of a few tenths of a picosecond per corner, which accumulates over a long run.

Tip 3: Ground, Where the Advice Is Inverted

Separating analog and digital ground is the one tip that should be removed rather than qualified. The reasoning behind it — keep noisy digital return currents out of sensitive analog areas — is sound; the implementation, a slot in the ground plane, is not. A return current flows in the plane directly beneath its trace, and at 10 GHz most of it is confined within about three dielectric heights: 0.76 mm on a 0.254 mm build. A slot under the trace forces that current to detour around the slot, which turns a narrow, well-controlled path into a loop of the slot’s length.

The working alternative is a continuous plane with functional partitioning. The analogue circuitry occupies one region, the digital another, and the signals are routed so that neither crosses into the other’s region. Where a crossing is unavoidable, the return path is provided locally — a stitching capacitor across the boundary, placed within a twentieth of a wavelength of the crossing — rather than by cutting the plane.

The corollary is that a ferrite bead joining two ground regions is not a fix. The bead and the plane capacitance on either side form a resonant tank whose impedance is lowest at resonance, which is precisely where the isolation was wanted. At a few megahertz a bead does something useful; above 100 MHz it is a liability, and above 1 GHz it is capacitive and does nothing at all. The same reasoning, with the numbers, is in EMC in layer-up design.

Tip 4: Loop Area, the One That Actually Radiates

Loop area survives every frequency change because it is the geometry of the radiating aperture. The far-field emission of a small loop rises with area, with current and with the square of frequency, so a loop that is 10 times larger in area radiates 20 dB more at the same current, and the same loop at 10 times the frequency radiates 40 dB more. This is why the same board that passes at 300 MHz fails at 3 GHz without any layout change.

LoopAreaRelative emissionTypical cause
Signal over continuous planeTrace width × dielectric heightReferenceCorrect routing
Signal with return 1 mm to the side≈ 10× reference+20 dBNeck-down, plane void, connector transition
Signal crossing a plane split≈ slot length × height+20 to +40 dBSplit ground, plane cut for a connector
Power feed without adjacent returnCentimetres of separation+30 to +50 dBPower routed as a single trace, return left to find its own way

The operational consequences are unglamorous. Ground stitching at λ/20 keeps the loop closed across a plane boundary. Return vias adjacent to signal vias keep the loop closed across a layer change. Keeping the power trace adjacent to its own return, or better, using a plane pair, keeps the supply loop small. None of these are routing aesthetics; each is a specific reduction in loop area of one to three orders of magnitude.

Tip 5: Vias, Right Conclusion for the Wrong Reason

“Minimise layer changes” is correct, and the usual justification — half a picofarad of parasitic capacitance per via — is the smaller half of the problem. Capacitance of 0.3–0.6 pF on a 50 Ω line produces a small impedance dip that is visible on a TDR and mostly harmless. The inductance, 0.5–1.2 nH per via depending on board thickness and barrel diameter, is what costs bandwidth, and the stub is what creates a resonance.

Board thicknessStub lengthQuarter-wave resonance in FR-4In PTFE
1.6 mm1.6 mm≈ 23 GHz≈ 32 GHz
2.4 mm2.4 mm≈ 15.6 GHz≈ 21 GHz
0.8 mm0.8 mm≈ 46 GHz≈ 63 GHz
Blind via, 0.9 mm0.9 mm≈ 42 GHz≈ 57 GHz

Read the table as a schedule: on a 1.6 mm FR-4 board, every through via is a resonator at 23 GHz, which is why millimetre-wave designs use thin laminates, blind vias or back-drilling rather than counting vias. The second-order effect of a layer change is the reference transition: the return current has to move to another plane, and without a return via the loop opens. One via per layer change is therefore the wrong unit of account; the correct unit is two vias per layer change, one signal and one return.

There is also a placement rule that saves both problems at once: put the return via within 0.5 mm of the signal via. That keeps the loop under a twentieth of a wavelength up to about 18 GHz and gives the returning current a path that does not have to find a decoupling capacitor. The geometry of the hole itself, including drill versus finished diameter, is in PCB through holes.

Tip 6: Decoupling Is Placement, Not Capacitance

“Place a decoupling capacitor near every power pin” is true and understated. What the capacitor does depends on the inductance between it and the pin, and that inductance is set by geometry: about 1 nH per millimetre of trace, 0.3–0.5 nH per via. A 0.1 µF part with a 0.9 nH connection resonates at about 17 MHz; the same part connected through 5 mm of trace and two long vias sees 4 nH and resonates at 8 MHz, which means it is inductive over most of the band it was chosen for.

The consequence for routing is specific: the capacitor should be placed so that its pads connect to the power and ground planes by vias, not by traces, and the via pair should be as close to the pads as the design rules allow. A capacitor placed 3 mm from the pin but connected by two short vias is better than one placed 0.5 mm away through a necked-down trace. Distance to the pin matters less than the loop inductance, and the loop includes the return path.

Above 100 MHz, no board-level capacitor helps much, because the plane pair and the package take over. The board’s job at those frequencies is to keep the loop small and the plane pair close — which is a stack-up decision rather than a routing one, and which is where the decoupling arithmetic in power supply noise on high-frequency PCBs starts.

What the Six Tips Leave Out

Three effects cause more failures on real high-frequency boards than any of the six tips, and none of them appears in the usual list.

The launch. A connector, a cable transition or an edge launch is a structure with its own impedance and its own cutoff frequency. An SMA connector is usable to about 18–26 GHz, a 2.4 mm connector to about 50 GHz, a 1.85 mm connector to about 67 GHz. A board designed to 40 GHz and fitted with SMA launches has a 40 GHz design that measures like an 18 GHz one, and no amount of routing discipline changes it.

The mask. Solder mask over an RF trace is a dielectric layer 20–30 µm thick with a Dk of 3.5–4.2 sitting on top of the line. It lowers impedance by 1–3 Ω and adds a lossy layer whose loss is rarely modelled. On a 50 Ω line that is a few percent; on a narrow high-Dk line it is more. The standard practice is to pull the mask back from RF traces and from ground contact areas, which is a fabrication note, not a layout preference.

Copper texture and glass weave. At 10 GHz the current is in the outer 0.65 µm of the foil, so the foil’s surface profile is part of the conductor. Standard electrodeposited foil at Rz 8–10 µm adds 20–45 % to conductor loss compared with HVLP foil at Rz ≤ 2 µm. And on glass-reinforced laminates, the difference in Dk between glass bundle and resin produces a periodic impedance variation along a trace, which appears as phase jitter on a differential pair — a few picoseconds, which at 25 Gbps is a meaningful fraction of the unit interval. Neither effect is visible in a routing review, and both are decided by material and copper callouts, as set out in high-frequency PCB stack-up design and in copper thickness in PCB.

The Same Six Tips at Three Frequencies

The clearest way to see which tips deserve effort is to score them by band. The matrix below is a judgement, not a measurement, but it encodes the arithmetic of the sections above.

Tip1 GHz10 GHz28 GHz
Short tracesHelpful (0.19 dB/100 mm)Critical (1.24 dB/100 mm)Dominant (2.72 dB/100 mm)
Avoid 90° cornersIrrelevantNegligibleMinor; use arcs for constant width
Split groundsHarmfulHarmfulHarmful
Loop areaImportantCriticalCritical
Fewer viasMildImportant (stub 23 GHz)Decisive (back-drill or thin core)
Decouple at the pinImportantImportant to 100 MHz onlyBoard-level decoupling irrelevant above 1 GHz

Three conclusions follow. Loop area is the only tip that is important at every frequency. Corner geometry is the only tip that is never important in the band where it is most often enforced. And the two tips that matter most above 20 GHz — via stub control and decoupling that acknowledges its own inductance — are the two that the usual one-line version of the advice states least precisely.

A Routing Review Checklist That Uses Numbers

1. Every RF net has a stated loss budget in dB, and the routed length is checked against it. 2. Every RF trace has a continuous reference plane, and no crossing of a plane boundary without a stitching path within λ/20. 3. Every layer change has a return via within 0.5 mm. 4. Stub length is stated, and back-drilling is called out where the stub would resonate in band. 5. Differential pairs are matched to a stated skew budget in picoseconds, not in mils. 6. Trace width is constant through bends and through the via field; necks are documented. 7. Mask pull-back is called out on RF traces and ground contact areas. 8. Decoupling capacitors are connected by via pairs, and the loop inductance is estimated rather than assumed. 9. The launch is specified by connector type and rated frequency. 10. The stack names the copper foil class, because at 10 GHz and above the foil is part of the conductor.

The list is longer than six items and shorter than a design guide. Its purpose is not completeness but checkability: each item can be answered with a number or a yes, which is what makes it survive a review with a fabricator who builds these boards every week.

Conclusion

The six tips for high frequency PCB routing are a reasonable starting set, provided each is attached to its number. Short traces are a per-length loss budget: 0.48 dB per 100 mm at 2.5 GHz, 1.24 at 10 GHz, 2.72 at 28 GHz on a 0.508 mm RO4003C microstrip. Loop area is the emission mechanism, and a ten-fold increase costs 20 dB. Vias are an inductance and a stub, not a capacitance: 0.5–1.2 nH each, resonating at 23 GHz for a 1.6 mm through via in FR-4. Decoupling is placement and loop inductance, 1 nH per millimetre, which is why 0.1 µF is useful at 17 MHz and not at 100 MHz.

Two of the six need correcting rather than qualifying. Splitting analog and digital grounds breaks the return path and should be replaced by functional partitioning on a continuous plane. And avoiding 90° corners, while harmless as a habit, addresses a discontinuity smaller than the etch tolerance below about 20 GHz; what actually matters at the bend is that the width stays constant.

XCE PCB routes and builds to these numbers: 2 to 72 layers, PTFE and hydrocarbon laminates, 0.127 mm cores, back-drilling and via fill, impedance held to ±10 % with coupon TDR, panels to 600×1200 mm, samples in as fast as 24 hours, inspected to IPC-A-600G and IPC-6015. The design sequence that precedes routing is in high-frequency PCB design steps, and the geometry of the lines being routed in microstrip and stripline design.

Frequently Asked Questions

Do 90-degree corners really cause signal reflections?
They cause a small one, and below about 20 GHz it is smaller than the discontinuity from the etch tolerance of the line itself. A right-angle bend adds a fraction of a picofarad, equivalent to a fraction of a millimetre of extra line. What causes measurable reflections at a bend is a change in trace width through the corner, which mitring does not fix but constant-width routing does.
Should analog and digital grounds be separated on a high-frequency board?
No. A split forces the return current to detour around the slot, and at 10 GHz the return current wants to stay within roughly three dielectric heights of its trace — 0.76 mm on a 0.254 mm build. Partition the functions by placement on one continuous plane, and where a crossing is unavoidable, provide a stitching capacitor within a twentieth of a wavelength.
How much does a via actually hurt a high-speed signal?
The capacitance is small, 0.3–0.6 pF, and mostly harmless. The inductance is 0.5–1.2 nH per via and the stub is the real problem: a 1.6 mm through via resonates at about 23 GHz in FR-4 and 32 GHz in PTFE, which is why millimetre-wave boards use thin cores, blind vias or back-drilling. Each layer change should also have a return via within 0.5 mm.
How close does a decoupling capacitor need to be?
Close enough that the connection inductance stays below about 1 nH, which in practice means via pairs at the pads rather than traces. A 0.1 µF capacitor with 0.9 nH resonates at 17 MHz; the same part with 4 nH of trace and via inductance resonates at 8 MHz and is inductive across most of its intended band. Loop inductance matters more than distance in millimetres.
How much loss does 100 mm of high-frequency trace cost?
On a 0.508 mm RO4003C microstrip with 50 Ω and 1 oz copper, about 0.48 dB at 2.5 GHz, 1.24 dB at 10 GHz and 2.72 dB at 28 GHz, including the dielectric and conductor terms. Halving the length halves those numbers, which is why layout compaction is the cheapest loss fix available before changing material.
What do the six tips leave out that matters most?
Three things: the launch (an SMA connector limits a board to about 18–26 GHz regardless of routing), the solder mask over RF traces (20–30 µm of Dk 3.5–4.2 material, worth 1–3 Ω), and copper texture (standard ED foil at Rz 8–10 µm adds 20–45 % to conductor loss against HVLP foil at Rz ≤ 2 µm). All three are material or fabrication decisions rather than routing ones.
RF & Microwave PCB Fabrication

Route It Once, With the Numbers Checked

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

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