Signal integrity is usually explained as a list of problems — reflection, crosstalk, skew, ground bounce — with a list of countermeasures beside it. That structure is easy to read and impossible to design against, because the five problems compete for the same budget and the countermeasures trade against each other. This is a detailed explanation of high-frequency PCB signal integrity in the form of five accounts: reflection, loss, crosstalk, skew and return-path noise, each with the formula or the table that sizes it, and then one worked budget that shows how they add up in a 10 Gbps link.
A detailed explanation of high-frequency PCB signal integrity has to start from what a receiver measures: an eye. The eye is closed by five mechanisms, and each one can be sized before the board is laid out.
| Account | What closes the eye | Sized by | Controlled by |
|---|---|---|---|
| Reflection | Ring and overshoot at discontinuities | Return loss against impedance tolerance | Stack, via design, termination |
| Loss | Rise-time degradation, vertical closure | dB per inch at the Nyquist frequency | Dielectric Df, copper roughness, length |
| Crosstalk | Noise added to the victim | Near-end crosstalk against s/h | Spacing, reference plane, parallel run |
| Skew | Horizontal closure | ps of mismatch against UI | Length matching, Dk uniformity, fibre weave |
| Return-path noise | Common-mode shift, jitter | L·di/dt across a broken plane | Plane continuity, stitching vias, layer changes |
The reason to frame it this way is that the five are not independent. Widening a trace to reduce loss changes the impedance unless the dielectric also changes; adding spacing to reduce crosstalk costs density; adding vias to stitch a plane adds reflections. A design that fixes one account by moving another has not improved the board, it has moved the failure.
Nothing else can be decided until the bandwidth is known, and bandwidth is set by the edge, not by the clock. The usual relation is BW ≈ 0.35/tr, where tr is the 10–90 % rise time of the driver.
| Edge (10–90 %) | Bandwidth | Nyquist of a 10 Gbps NRZ link | What the laminate has to be |
|---|---|---|---|
| 1 ns | 350 MHz | — | FR-4 is fine; loss is irrelevant at these lengths |
| 300 ps | 1.2 GHz | — | FR-4 acceptable up to about 300 mm |
| 100 ps | 3.5 GHz | 5 GHz | Mid-loss laminate or short FR-4 runs |
| 50 ps | 7 GHz | 5 GHz | Low-loss laminate; copper roughness now matters |
| 25 ps | 14 GHz | 5 GHz | Ultra-low-loss; the package and connector dominate |
The right-hand column is where most design reviews go wrong: they choose the laminate from the data rate and ignore the driver. A 10 Gbps link with a 100 ps edge and a 50 mm channel runs comfortably on a mid-loss laminate, while the same data rate with a 25 ps edge and a 400 mm backplane channel may not close on the best laminate available.
Every impedance discontinuity reflects a fraction of the incident wave, and the fraction is set by the two impedances: Γ = (Z2 − Z1)/(Z2 + Z1). The consequence is that a tolerance on impedance is a tolerance on return loss, and the conversion is not linear.
| Impedance deviation | Reflection coefficient | Return loss | Visible effect on a 100 ps edge |
|---|---|---|---|
| 2 % | 0.010 | −40 dB | Not measurable |
| 5 % | 0.024 | −32 dB | Negligible |
| 10 % | 0.048 | −26 dB | A few percent of eye height |
| 20 % | 0.091 | −21 dB | Clear ring; jitter increases |
| 50 % | 0.200 | −14 dB | Eye closure that no equaliser recovers |
Which is why ±10 % impedance is a usable specification and ±5 % is an expensive one: the return loss improves by 6 dB for a halving of the tolerance, and 6 dB of return loss buys very little eye at 10 Gbps unless the line is short. It is also why a single 50 % discontinuity — a connector, an unoptimised via, an antipad that is too large — dominates everything else on the channel. The design side of the same number, including the fab tolerance budget that produces it, is in impedance control in PCB manufacturing.
Insertion loss is the sum of a dielectric term that scales with frequency and a copper term that scales with the square root of frequency, and the split between them decides which material change is worth making.
| Laminate | Df | Dielectric loss at 10 GHz | Total with 1 oz copper | Coin |
|---|---|---|---|---|
| FR-4 | 0.020 | 1.08 dB/in | 1.28 dB/in | Dielectric is 85 % of the total |
| RO4350B | 0.0037 | 0.18 dB/in | 0.38 dB/in | Roughly even split |
| RO4003C | 0.0027 | 0.13 dB/in | 0.32 dB/in | Copper now dominates |
| RT/duroid 5880 | 0.0009 | 0.035 dB/in | 0.22 dB/in | Copper is 84 % of the total |
That flip is the single most useful fact in high-frequency signal integrity. On FR-4 the dielectric is the problem, so a better laminate fixes almost everything. On a 0.0009 laminate the copper is the problem, so a better laminate fixes almost nothing — what fixes it is smoother foil, a wider line, or a shorter path. This is why “use a lower-loss material” is only an answer when the material is currently the larger term, and why a design moving from FR-4 to 4350B gains 0.9 dB/in while a design moving from 4350B to 5880 gains 0.16 dB/in for several times the material cost.
The formula behind the dielectric column is close to αd ≈ 2.6·fGHz·√Dk·Df dB/in, which is worth keeping because it shows that dielectric loss depends on the product of Dk and Df, not on Df alone. A high-Dk material can lose more than a low-Dk one at the same Df.
Coupled noise falls steeply with the gap between traces relative to the dielectric height, and it also depends on how long the two traces run parallel.
| Gap / height | Near-end crosstalk | Parallel run that keeps it under 5 % of the victim swing |
|---|---|---|
| 1 | −6 dB | Any length: the coupling is too strong to design around |
| 2 | −14 dB | Under about 10 mm |
| 3 | −20 dB | Under about 30 mm |
| 4 | −26 dB | Under about 60 mm |
| 6 | −34 dB | Effectively unlimited for most designs |
Two corrections are worth stating because both appear in most explanations. A wider trace does not reduce crosstalk — at a fixed gap it couples more, since the facing area grows. And the digital 3W rule is not the RF 3h rule: 3W measures centre-to-centre in units of trace width, 3h measures edge-to-edge in units of dielectric height, and on a 0.254 mm stack with a 0.59 mm line they differ by 55 %. The full comparison is in trace width and spacing.
Propagation delay is set by the effective Dk, so a length mismatch converts directly into a timing error. On a microstrip with εeff ≈ 3.3 the delay is 6.1 ps/mm (154 ps/inch); on a stripline in the same laminate it is about 5.8 ps/mm because the field is entirely in the dielectric.
| Allowed skew | Length mismatch, microstrip | Length mismatch, stripline | Where it applies |
|---|---|---|---|
| 5 ps | 0.8 mm | 0.9 mm | 25 Gbps lanes with no per-lane deskew |
| 10 ps | 1.6 mm | 1.7 mm | 10 Gbps differential pairs |
| 25 ps | 4.1 mm | 4.3 mm | 25 % of a 100 ps unit interval |
| 100 ps | 16 mm | 17 mm | Parallel buses with a common clock |
Intra-pair skew is only one of the three sources. Inter-pair skew matters as much on a wide bus, and it is usually caused by routing topology rather than by length. The third source is material: glass weave in a laminate makes Dk vary by a few percent from point to point, which at 6 ps/mm is a few picoseconds across a board, and is why very high-speed designs use spread glass or route at an angle. None of the three is fixed by a length-tuning tool alone.
A signal current is half of a loop. The return current follows the reference plane directly beneath the trace at low frequency, and as frequency rises it migrates to the path of least impedance — which is still directly beneath the trace, but confined to a narrow band. The crossover frequency between resistance-dominated and inductance-dominated return is f = R/(2πL), which for a 1 oz plane is in the low megahertz — the physics is worked through in concepts of electricity in PCB.
| Condition | What the return does | Measured consequence |
|---|---|---|
| Continuous plane | Flows directly under the trace, within about 3h | Loop area minimal; radiation and crosstalk low |
| Slot in the plane | Detours around the slot, forming a loop | Radiated emission and crosstalk rise by tens of dB at the slot length’s resonance |
| Layer change at a via | Has to jump planes through the nearest stitching via | Without a stitching via within a few millimetres, the loop includes the whole plane pair |
| Ground bounce | Plane inductance shared by many drivers | 0.5 nH at 0.5 A/500 ps is 0.5 V of common-mode shift |
The ground-bounce row is where a common countermeasure fails: decoupling capacitors do not fix ground bounce. A capacitor supplies charge for a slow current demand; bounce is L·di/dt across the plane inductance at the edge rate, and the fix is more parallel return paths — more stitching vias, a continuous plane, and fewer simultaneous switching outputs sharing one return.
A via is a short cylinder of copper with a capacitance to the planes it passes and an inductance along its length, and the two form a discontinuity whose severity depends on the antipad and the stub.
| Stub length | First λ/4 resonance (FR-4) | What it does to the eye |
|---|---|---|
| 1.6 mm (through via, no back-drill) | About 23 GHz | Notch at 23 GHz, and rolling attenuation from a third of that |
| 1.2 mm | About 31 GHz | Acceptable to about 15 GHz |
| 0.5 mm (back-drilled) | About 75 GHz | Effectively transparent to 25 GHz |
| 0.08 mm (HDI microvia) | Above 400 GHz | The microvia is not the limit; the antipad is |
Three controls follow. Back-drill through vias that carry signals above about 10 GHz, to a residual stub of 0.2–0.3 mm. Size the antipad for the impedance rather than for the drill, because an oversized antipad is a capacitance to the plane and an undersized one is a plating risk. And treat a via field as a distributed discontinuity rather than a set of individual ones: fifteen vias in a row at 1 mm pitch is a periodic structure, and it has a stopband.
The structure of the via itself, and the aspect-ratio rules that decide whether it can be plated at all, are in PCB through hole.
A worked example makes the trade-offs visible: a 10 Gbps NRZ link, 100 ps edge, 150 mm of microstrip on a mid-loss laminate, one connector, two vias, a 50 mm parallel run at s = 3h.
| Account | Estimate | Share of the eye |
|---|---|---|
| Loss | 0.38 dB/in × 5.9 in ≈ 2.2 dB, plus 0.5 dB connector | Rise time 100 ps → about 130 ps; the largest single term |
| Reflection | Two vias and one connector, each ≤10 % discontinuity | A few percent of height, plus some deterministic jitter |
| Crosstalk | −20 dB coupled over 50 mm | About 3 % of the swing on the victim |
| Skew | 1 mm intra-pair mismatch = 6 ps | 6 % of the 100 ps UI |
| Return path | Two layer changes, stitching vias within 2 mm | Negligible if stitched; dominant if not |
The budget says what the layout review should focus on: loss and return path. Loss is set by length and laminate before routing starts, and the return path is set by via placement, which is decided in the first hour. Reflection and crosstalk are second-order here, and skew is a routing constraint that a length-tuning pass can meet. A team that spends the review arguing about crosstalk and ships with a plane split under the connector has optimised the wrong account.
The stack is the only place where all five accounts can be set at once, and there are four rules that do most of the work. Put every signal layer next to a continuous plane, so reflection and return path are both defined. Keep the plane pair used for the highest-speed signals adjacent, so the stitching vias are short. Route the fast signals on the layers closest to the outside, where the dielectric is thinnest and the vias are shortest. And keep the power plane out of the signal reference path — a signal referencing a split power plane is a signal referencing a slot.
For a high-frequency design the last rule usually forces a decision between layer count and stack symmetry. A 6-layer stack can reference every signal to ground, but it will have only four routing layers; an 8-layer stack gives two more routing layers for one more press cycle and tighter registration. The trade, and the plane-counting rule behind it, is in high-frequency PCB stack-up design.
| Statement | What actually happens |
|---|---|
| Use a low-loss laminate such as enhanced FR-4 | Enhanced FR-4 has a Df around 0.010–0.015, which is still five to fifteen times 4350B. It is better than standard FR-4 and it is not low loss; at 10 GHz it is still about 0.7 dB/in of dielectric loss |
| Decoupling capacitors fix ground bounce | They supply charge for slow current demand. Bounce is L·di/dt at the edge rate, and the fix is more parallel return paths and fewer simultaneous switching outputs sharing one return |
| Keep spacing at three times the trace width | That is the digital 3W rule, measured centre-to-centre. RF practice measures edge-to-edge against dielectric height (3h), and the two differ by more than 50 % on a typical stack |
| Skew only matters for differential pairs | A parallel bus with a common clock has a window as well, and inter-pair skew from routing topology is often worse than intra-pair skew from length |
| A lower-Df material always improves the eye | Only while dielectric loss is the larger term. On a 0.0009 laminate copper is 84 % of the loss, so the next material change buys 0.16 dB/in while a smoother foil or a wider line buys more |
Three measurements cover the five accounts. A TDR on a coupon gives impedance and reflection, and its resolution is set by the edge: a 35 ps step resolves about 3 mm of trace, so a coupon shorter than that cannot show a local discontinuity. A VNA gives insertion loss and return loss against frequency, including the via resonance. And an eye diagram at the receiver gives the sum of all five, which is the only number the customer cares about.
The trap is measuring only one of them. A coupon that passes impedance and a line that passes loss can still fail the eye, because the eye is also paying for crosstalk, skew and return path. The test ladder and what each step can and cannot prove is set out in high-frequency PCB testing.
Signal integrity on a high-frequency board is five accounts, not five problems. Reflection converts an impedance tolerance into a return loss, and 10 % gives −26 dB while 50 % gives −14 dB. Loss splits into a dielectric term that follows Df·√Dk and a copper term that follows roughness and width, and the split decides which material change is worth buying: on FR-4 the dielectric is 85 % of the loss, on a 0.0009 laminate the copper is 84 %. Crosstalk is a spacing ratio, about −20 dB at s = 3h. Skew is time priced in millimetres, at 6.1 ps/mm on a microstrip. And the return path is where most failures actually happen, because it is the only account that is decided by via placement rather than by the drawing.
Add them on one budget before the layout starts, and the review has something to argue about that can be measured. The other half of the same physics — what happens to the heat those losses become — is in the companion article on high-frequency PCB thermal management.
XCE PCB builds the stack and measures the result: 2 to 72 layers, impedance ±10 % verified on coupons from the centre and the edge of every panel, back-drilling to a 0.25 mm residual stub, 600×1200 mm panels, samples in as fast as 24 hours, inspected to IPC-A-600G and IPC-6015. Send the stack and the eye mask together and the budget can be closed before the first panel is cut — request a quote or WhatsApp +86 180 3341 4657.
Send us your Gerber files and requirements for a manufacturability review and a tailored quotation.