Bottom line first: the fiber weave effect is not a laminate defect — it is the inevitable result of a routing direction that does not match the warp and weft geometry of the glass cloth. Of two lines in the same differential pair, one sits on a glass bundle and the other falls in a resin-rich "window", so their delay per unit length differs; the longer the link, the larger the accumulated phase skew. At 112G and above the intra-pair skew budget is often only a few picoseconds, and length tuning at the end cannot buy it back — it has to be solved at stackup and routing-rule stage.
Turning the "window" into a quantifiable delay difference
Standard E-glass cloth has a dielectric constant of about 6.0–6.6, and epoxy resin about 3.0–3.3. After weaving, the local effective Dk on a bundle and in a window can differ by 0.15–0.3, which corresponds to a propagation-velocity variation of roughly 3%–6%: a 100 mm trace that lands in the two extreme positions can show a delay difference of 5–20 ps.
- The unit interval of 112G PAM4 is about 17.9 ps, and of 56G PAM4 about 35.7 ps.
- Most SerDes specifications set the intra-pair skew budget within a channel at 3–8 ps, and the share left for the PCB is often only 2–4 ps.
- In other words, a run on the order of 10 mm sitting on a bundle is already enough to consume the budget.
Four control methods and what each costs
The signature of this effect is unusual: it is a fixed skew strongly correlated with trace length and routing direction, not random jitter. On the eye diagram it appears as shifted edge positions and increased mode conversion, and the bit error rate degrades as the rate rises. Skew can differ between lots of the same stackup and even between layers of the same board, which is why it is so often misdiagnosed as a device problem during debug. A single skew measurement on the same layer and the same dielectric usually saves a round of unnecessary device swaps.
- Spread or flat glass: replacing twisted bundles with dispersed filaments or flat bundles shrinks the window to the order of a few microns and gives the best Dk uniformity, at the cost of material price and lead time. The next best option is thin glass such as 1035 or 1037 used in two plies, which averages the weight of the local non-uniformity.
- Angled routing: running traces at 5°–15° to the warp avoids long parallel runs along a bundle. Suppression is clear above 8°, but it adds roughly 1%–3% of extra trace length and complicates BGA escape; volume practice commonly takes 5°–10°.
- Intra-pair phase compensation: small serpentine or bump compensation in the breakout area. Do not concentrate the compensation in one place, or it introduces a new impedance discontinuity.
- Stackup adjustment: replacing a single ply of 2116 or 7628 with two plies of 1080, or moderately increasing dielectric thickness, reduces the weight of local Dk variation on propagation velocity — but it moves the impedance and loss targets with it.
Three data sets to request at acceptance
- Material list: glass style, whether it is spread, and measured Dk and Df with the test method (SPDR or cavity resonator preferred) — not just the nominal values from the datasheet.
- Skew measurement: TDR, TDT or phase measurement sampled across the pair; high-speed links should be held within 3 ps, and general high-speed signals within 5 ps.
- Impedance and line width: impedance tolerance ±8% (against the usual ±10%) and line-width accuracy ±8 μm, both reported in the same document, because line-width variation adds to Dk variation rather than averaging out.
Four items to write into the RFQ
- The rate class and protocol, for example 112G PAM4 or PCIe 6.0, since these set the skew budget.
- The longest differential-pair length and the number of pairs per group.
- Whether a glass specification is mandatory, or whether an "equivalent performance" substitution is allowed.
- Whether a small line-width adjustment for better uniformity is acceptable, given that it shifts the impedance target.
Our stackup range covers 1–44 layers, finished copper up to 13oz, a maximum panel size of 610×1100 mm and aspect ratios up to 20:1 (25:1 for backplanes), with 24-hour expedited prototyping. Raising the rate class, the stackup and the impedance requirement in the same engineering review saves a round of correspondence.
Send your Gerber files and stackup requirements; a 24-hour expedite channel is available.