Contact Us
  • Home
  • BLOG
  • Why Does Glass Weave Skew a Differential Pair?

Why Does Glass Weave Skew a Differential Pair?

A differential pair picks up skew when its two traces travel at different speeds, and on woven glass the speed difference comes from what each trace sits over. A trace that runs along a glass bundle sees mostly glass, with a dielectric constant near 6.1 to 6.3 for E-glass at 1 GHz. A trace that runs over a resin window between bundles sees resin, near 3.2 to 3.4. When one line of the pair is on a bundle and the other is on a window, the two leave the driver with different delays and the gap grows with every millimetre of parallel run.

The size of the effect is easy to put on paper. Propagation delay is close to 85 times the square root of effective dielectric constant, in picoseconds per inch. At an effective Dk of 4.0 that is 170 ps per inch, near 6.7 ps per mm. If one line sees 4.0 and the other sees 4.4, the delays are 170 and 178 ps per inch, so the pair gains about 8 ps of skew per inch of length. Over 250 mm of parallel routing that is near 80 ps, and at 25 Gbps, where one unit interval is 40 ps, that is two unit intervals. The link is out of spec long before anything runs warm.

Why the weave only bites at some trace widths

Bundle pitch in a standard fabric is a few tenths of a millimetre, and bundle width is close to the width of a 0.15 to 0.25 mm trace. A trace narrower than the bundle tends to sit either on glass or off it, which is the worst case for velocity difference. A trace much wider than the bundle averages glass and resin across its width, so its effective Dk lands nearer the mean and the skew falls. That is why the same stack and the same serialiser can be clean on one board and marginal on another: the pair geometry against the fabric changed, not the laminate.

Resin content and glass style move the same ratio. A high resin content laminate spreads the bundles apart and widens the resin window between them. Spread and flat glass fabrics pack more filaments into a flatter bundle, so the trace sees a more even medium. Standard 1080 and 2116 fabrics have a coarser bundle pitch against trace width than 1035 or 1067, so a 0.1 mm pair on a coarse fabric is more exposed to the effect, and glass with a lower Dk changes the absolute number without removing the difference.

What shows on the bench

Weave skew hides from the checks that usually get run. It does not break continuity, it does not change a length measurement, and a 1 MHz capacitance check does not see it. Two traces matched to 0.1 mm can still differ by 60 ps, because the difference is in velocity, not in length.

It turns up in three places. First, in intra-pair skew measured by TDR, or by phase on a network analyser at the operating frequency. Second, in common mode conversion: the pair stops being balanced and part of the differential signal becomes a common mode component that radiates and that the receiver rejects poorly. Third, in the eye at the far end of a long backplane, where a card that passed on a short coupon closes at 60 to 80 cm of trace.

Where builds go wrong

Matching length in the CAD tool and calling it done. Equal length is necessary and not sufficient. A pair matched to 0.05 mm but crossing bundles at a different rate on its two lines still carries tens of picoseconds. The check that matters is skew at frequency or in time on the finished stack, not the length report.

Qualifying on a coupon routed at a different angle. Coupons usually run straight along one axis. A coupon routed at 10 to 15 degrees to the fabric axis averages the weave, so it looks better than the production route, which may run parallel to a bundle. When the board and the coupon disagree, the coupon is the one that is wrong.

Changing fabric or laminate for the same Dk on the datasheet. Two laminates with the same nominal Dk can carry a different bundle pitch and different resin content, and that is the variable that sets skew. When the stack changes, the pair has to be measured again, even if the impedance still comes out at 100 ohms.

How we hold it on the panel

At SUPERB PCB the high-speed stack is quoted with the fabric named, not only the laminate family, and the routing angle against the fabric axis is checked at CAM review before the first panel is tooled. Where a customer pair is narrow against a coarse fabric we quote a spread glass option and say what it does to Dk and to lead time. Impedance is held at plus or minus 5 percent, and to plus or minus 2 percent on critical nets, with a TDR coupon on every panel: differential impedance, intra-pair skew and insertion loss go on the same coupon so one measurement cannot contradict the other.

Lines down to 0.05 mm are available on high layer count and HDI work, back drilling is quoted to a stub of 0.25 mm or less where a via stub would dominate the loss, and the same laminate family is held across builds so a re-qualification does not start from zero. The material and stack discussion sits on high-speed PCB, the digital side on high-speed digital capability, and where roughness is the limit rather than weave, in copper foil roughness above 20 GHz.

What to send with the question

The stack-up with fabric and resin content, the pair width and gap, the length of the parallel run, the data rate and the coding, the skew the serialiser can absorb, and the measurement you will use to accept the board. With those six the answer is a fabric, a routing angle and a coupon specification, instead of a hope that the length match covers it.

Need a quote or a review of your stack against these limits? pcba@superb-tech.com · WhatsApp +86 133 9608 1443