PCB Differential Signaling: Impedance and Length Matching
Differential signaling sends one signal down two traces instead of one and recovers the data from the voltage between them rather than from the voltage to ground. That one change buys immunity to a large class of interference, but it also imposes rules that a single ended design never has to think about. A pair that is laid out carelessly can be worse than a single trace, so the geometry has to be treated as part of the electrical design rather than as a routing detail.
What the Pair Does That a Single Trace Cannot
The two conductors carry equal and opposite currents, and the receiver looks only at the difference between them. Any disturbance that affects both traces equally appears as a common mode signal and is rejected by the input stage, which is why differential signaling remains usable in an environment where a single ended line would be swamped by switching noise from motors, relays or a nearby power stage.
The second benefit is radiated. Two currents flowing in opposite directions in close proximity produce magnetic fields that largely cancel, so the pair radiates less than a single trace carrying the same data rate. That is a large part of why every modern high speed interface, from USB and Ethernet to PCIe and automotive networks, is defined as a differential pair.
Where the Characteristic Impedance Comes From
Differential impedance is not simply twice the single ended value, because the traces couple to each other as well as to the reference plane. Bringing them closer reduces the impedance, raising them or thinning them increases it. Interface standards fix the target: ninety ohms for USB, one hundred ohms for Ethernet and HDMI, eighty five ohms for PCIe, with tolerances that are often plus or minus ten percent.
The number has to be produced by geometry, not by hope. Trace width, spacing, dielectric thickness and dielectric constant all enter the calculation, and the same geometry will not give the same result in microstrip and in stripline because the field distribution differs. The structures and their trade offs are described in microstrip and stripline routing.
Length Matching and Skew
If the two traces differ in length, the two signals arrive at different times and part of the differential signal converts into common mode. The usual rule is to match within about five thousandths of an inch for general high speed work and within two or three for the fastest interfaces, but the tolerance should come from the standard rather than from habit.
Matching has to be done without destroying the coupling. The shorter trace is lengthened with a detour, and the detour should be smooth and long rather than a tight accordion, because tight serpentine turns produce their own impedance discontinuities and can radiate. Pair to pair matching matters as well when several lanes must arrive together, as in a multi lane bus.
Vias, Reference Planes and Return Paths
A via changes the impedance of the pair, adds a stub that behaves as a resonant element at high frequency, and creates an opportunity for the two traces to be treated differently if their transitions are not identical. Both traces of a pair should change layer at the same place, with ground vias close by to give the return current a short path, and thick boards often need back drilling to remove the stub.
The reference plane matters just as much. A pair that crosses a plane split forces its return current to take a long detour, which raises impedance and radiates. Keep each pair over one continuous reference, and if a layer change is unavoidable, change to a plane of the same net. The broader discipline is described in high frequency trace and data bus routing.
Common Mode Noise and How It Appears
Dominant mode conversion is the enemy. A pair in which one trace has a longer stub, a different pad shape or a slightly different distance to ground will convert part of the differential signal into common mode noise, and common mode current is what radiates from cables and fails an EMC test. The layout should be as symmetric as the design allows, and symmetry should be checked rather than assumed.
Where a design still produces more common mode than the standard allows, a common mode choke at the connector is the usual remedy. It presents high impedance to the common mode component while passing the differential signal almost unchanged, and it is far easier to place a component than to rework a tightly routed pair on an already fabricated board.
Materials, Loss and Skew
Signal integrity over distance is limited by loss, and loss rises with frequency. Standard FR4 is adequate for short runs and moderate rates, while long backplanes and the fastest serial links call for a laminate with a lower dissipation factor. The choice is not about prestige: a lossier material closes the eye diagram and forces retimers or a shorter channel.
Fiber weave adds a second effect. Because glass bundles and resin have different dielectric constants, two traces a few millimetres apart can see different effective dielectric constants, which introduces skew within a pair. Rotating the routing angle relative to the weave, or specifying a spread glass laminate, reduces the problem on very long high speed channels.
Loss is also a matter of geometry. A wider trace has lower resistive loss, but widening the pair changes the impedance, so width and spacing have to be adjusted together rather than one at a time. An inner layer gives a controlled dielectric environment, while an outer layer exposes the pair to solder mask and to whatever the enclosure places above it. These decisions are made when the stackup is defined, and they are expensive to revisit after the design is released.
Fabrication Tolerance and Verification
Etch factor, layer to layer registration, dielectric thickness and copper thickness all vary in production, and each of them moves the impedance. A fabricator that offers controlled impedance measures a test coupon on the panel with time domain reflectometry and adjusts the process, which is why the impedance requirement belongs on the fabrication drawing together with the tolerance and the reference structure.
Ten percent is a reasonable general tolerance and five percent is available for a premium. The design should also state which layer and which structure the coupon represents, because a coupon that does not correspond to the real stackup proves nothing. The practical sequence is to agree the stackup and the impedance targets with the fabricator before routing begins, and then to route to the geometry that the stackup implies. If the board later needs a different laminate or a different layer count, the impedance calculation has to be redone and the routing revisited, which is why that conversation belongs at the start of the project rather than at the quotation stage. Where the pair runs past other fast nets, keep the separation that the crosstalk budget requires, as described in the 3W rule and crosstalk control.
FAQ
Can a differential pair be routed on two different layers? It can, but each change of layer adds vias and stubs that disturb the pair. If a change is unavoidable, make it for both traces at the same point and provide return vias beside them.
Does a differential pair need a ground plane? Yes. The pair needs a continuous reference for its return current and for the impedance calculation to hold. Removing the plane under a pair changes the impedance and increases radiation.
How tight should length matching be? Follow the interface standard. General high speed designs are usually matched within five thousandths of an inch, and the fastest serial links within two or three.
