A 3 mm move worth 35 dB

On a SerDes board failing radiated emissions at 1 GHz, the entire fix was relocating one 2.2 µF X7R ceramic 3 mm closer to the IC power pin and dropping its via pair onto the pad. Measured improvement at 1 GHz: 35 dB. No new components, no stackup change. The λ/20 placement rule predicted none of this — because λ/20 is the wrong physics for decoupling.

Why λ/20 is the wrong model

The λ/20 rule treats the cap-to-pin connection as a transmission line and asks whether it is "electrically short." At 1 GHz in FR-4, λ/20 ≈ 7.5 mm, so a 3 mm move looks negligible. But a decoupling capacitor's job above ~100 MHz is not set by trace length — it is set by the loop inductance of the current path through the capacitor's mounting.

A real capacitor is a series RLC: capacitance C, equivalent series resistance ESR, and a mounted inductance L_mount that includes the part's own ESL plus the pads, traces, and — dominantly — the vias connecting it to the planes. Its impedance is:

Z(ω) = √(ESR² + (ωL_mount − 1/(ωC))²)

The self-resonant frequency is f_SRR = 1/(2π·√(L_mount·C)). Above f_SRR the capacitor is inductive — it no longer decouples; it is just an inductor whose value is L_mount. For a 2.2 µF X7R with ~1.5 nH part ESL, f_SRR is near 3 MHz. At 1 GHz, that capacitor is entirely L_mount, and:

Z(1 GHz) = 2π·(1×10⁹)·L_mount

Every picohenry of mounting inductance matters. A standard via is ~0.3–0.5 nH; a 1.6 mm-deep via pair adds ~0.7 nH; 3 mm of 0.2 mm trace adds ~1.8 nH. Move the cap 3 mm closer and shorten the via drop, and L_mount falls from ~3.2 nH to ~1.4 nH.

⚠️Distance is a proxy — inductance is the cause

Do not specify "place within X mm." Specify a mounted-inductance budget (e.g., L_mount ≤ 1.5 nH) and a via strategy (via-in-pad, double vias, minimum drill-to-plane depth). Two boards with identical cap-to-pin distance can differ 6 dB if one uses a single via and the other uses a via pair in-pad.

The data

Measured impedance at 1 GHz for a 2.2 µF X7R, varying the via/trace path length:

Each millimeter of added separation costs 3–4 dB of conducted/radiated performance above 800 MHz, and the relationship is exponential in effect because Z is linear in L and L grows with both trace length and via count.

The method: target impedance

Stop placing by distance. Place by target impedance Z_target — the maximum power-rail impedance that keeps voltage ripple within budget for the load's transient current:

Z_target = ΔV_allowed / ΔI_transient

For a 1.0 V rail, 5% ripple budget, 4 A transient step: Z_target = 0.05 V / 4 A = 12.5 mΩ — at low frequency. At 1 GHz the realistic goal is the < 1.5 Ω mounted-impedance threshold above, met by the on-die and on-package capacitance plus the closest board cap with minimum L_mount. The board cap's only job at 1 GHz is to keep L_mount low; choose the part for low ESL, then design the mounting to win.

Also: a ground-plane discontinuity > 0.5 mm in the return path forces return current to detour, adding loop inductance and re-radiating. Discontinuities of this kind correlate with the majority of high-frequency emissions failures. Probe the actual return path, not the schematic net.

POCONS connection

When the switching node's near-field still couples to a cable or aperture after the decoupling is optimal, the residual is a radiated problem, not a PDN problem. A POCONS board-level shield over the regulator and its decoupling array contains the near field at the source — the complement to a clean PDN, not a substitute for it.

One thing

The λ/20 rule asks "is the wire short?" The right question is "how much inductance is in the loop?" Budget L_mount, use via-in-pad, and probe the real return path. A 3 mm move can be worth 35 dB because 3 mm of trace plus an extra via is 1.8 nH — and at 1 GHz, 1.8 nH is 11 Ω.