Four routes by which a disturbance on a supply rail becomes an error at a receiver — and the observation that two of the four are not power problems at all but return-path problems wearing a different hat.
Power integrity earns its place in a signal-integrity series only through the routes by which it damages a signal. There are four worth separating, and they are not equally famous.
Rail noise modulates the output stage's supply and therefore its swing and its threshold, producing amplitude noise directly — which deck 10 showed is also jitter, through the slew rate.
Rail noise reaches the phase-locked loop's control node and modulates its frequency, which integrates to phase and appears as jitter. This is the best-known route and the subject of slides 03 and 04.
A signal that changes reference plane needs its return current to cross between two planes, which means through the power delivery network. The network's impedance is therefore in series with the signal. Slides 05 and 06.
Every via passing through a plane pair couples into it, and the pair is a shared resonant medium. What one signal injects, another picks up. Slide 07.
The third and fourth are the ones most often missed, and they are not really power problems. They are the return-path argument of deck 02 restated: the power delivery network is the return path whenever a signal changes reference, so its impedance becomes a signal-integrity parameter whether or not anyone intended it to.
When many drivers switch at once their currents all flow through whatever inductance the supply and return paths share, and the voltage that develops across it is $L$ times the total rate of change. The scaling with driver count is what makes this an interface problem rather than a component one.
A differential pair draws a constant current from the supply whichever way it is driven, because one leg sources exactly what the other sinks. Its contribution to this sum is therefore close to zero. Deck 05 listed the benefits of differential signalling and this one rarely appears on such lists; on a wide interface it is frequently the largest of them.
A tone on the supply reaches the oscillator's control node attenuated by whatever rejection exists, and modulates its frequency. Frequency modulation integrates to phase modulation, so the phase deviation is the frequency deviation divided by the modulation frequency — and the jitter is that phase divided by the carrier's angular frequency.
The loop then filters. Against a disturbance injected at the oscillator its error response is high-pass, and how steeply it rises decides the whole answer: a first-order loop rejects in proportion to frequency, a second-order loop in proportion to frequency squared. Since the modulation index is already falling as the reciprocal of frequency, the first-order loop's rising rejection cancels it exactly and the jitter is flat from the lowest frequencies up to the loop bandwidth. Only in the second-order case does the square win, and only then is there a genuine worst frequency.
Both are plotted, because the familiar rule — that the dangerous tone is the one near the loop bandwidth — is a statement about second-order loops and is simply false for a first-order one. Real synthesis and clock-recovery loops are type II, so the rule survives; but it survives for a reason worth knowing, and the flat curve is what shows it.
A switching regulator puts a large, narrow, entirely predictable tone on the rail at its switching frequency. The previous slide says what that tone becomes, and therefore gives a design rule that is easy to state and routinely ignored: the regulator's frequency should be chosen against the transceiver's loop bandwidth, not against efficiency alone.
The loop tracks and corrects it, and the correction improves faster than the modulation index grows, so the jitter falls away at twenty decibels per decade below the natural frequency. A hundred-kilohertz regulator under a four-megahertz loop is thirty decibels better placed than a four-megahertz one, which is why low-frequency regulators were never a jitter problem.
The worst place, and where modern multi-megahertz regulators land by default. The response is symmetric about the loop's natural frequency — twenty decibels per decade up on one side, twenty down on the other — so “just above” and “just below” are equally bad and the peak itself sits at the bandwidth, not beyond it.
Also acceptable: the phase deviation falls inversely with the modulation frequency even though the loop does nothing. Very high switching frequencies are benign for this reason, whatever else they cost.
Deliberately modulating the switching frequency smears the tone across a band. That reduces the peak and leaves the total energy where it was, which helps an emissions measurement and helps a jitter budget rather less than is usually assumed.
Deck 02 treated this as a return-path question and priced it as a loop inductance. Smith and Bogatin treat the same configuration as a power-integrity question, and the change of view is worth adopting because it explains the remedy.
When a signal via carries a trace from a layer referenced to ground to one referenced to a supply, the return current has to cross from one plane to the other. The only route is the cavity between them, so the cavity's impedance appears in series with the signal. Everything deck 12 established about that impedance — that it resonates, that it depends on where you probe, that its peaks are tall — is therefore a signal-integrity property.
At a cavity resonance the impedance the return current sees is at its highest, so the signal suffers the largest series discontinuity at exactly the frequencies where the plane pair rings. A transition that is acceptable at most frequencies can be severe at a few.
That the cheapest fix is to not do it: plan the layer transitions so that a high-speed net either keeps its reference or changes between two planes of the same net, which needs only a stitching via. This is a stackup decision made before routing, and it cannot be retrofitted.
Where the transition is unavoidable, the return current crosses through capacitors placed between the two planes. The question is how many, and it has the same answer as the one in deck 11: above their resonances, capacitors present mounting inductance, so the count is set by inductance rather than by capacitance.
Blocking capacitors placed between planes do two different things, and the number needed is not the same for each. Carrying the return current of a particular transition needs a capacitor close to that transition. Damping the cavity's resonances needs capacitors distributed across the plane pair, in a number set by the cavity rather than by any one signal. A design that provides for one and not the other usually discovers it in an emissions chamber.
The fourth route is the least intuitive. A plane pair is a two-dimensional waveguide shared by everything that penetrates it. A via injecting energy into it does not deliver that energy to a particular destination; it excites the cavity's modes, and every other via in the same cavity is coupled to those modes.
That makes the plane pair a crosstalk mechanism with a quite different character from the one in deck 06. Trace-to-trace coupling falls off with separation; cavity coupling does not, because a resonant mode fills the whole cavity. Two vias at opposite corners of a board are coupled through a mode that spans it, and moving them apart does not help.
Lowering the quality factor of the modes reduces the coupling at every location simultaneously, which spacing cannot do. Distributed capacitors and deliberately lossy planes both work.
Closer planes lower the cavity's impedance in proportion, so the same injected current produces less voltage. The same lever as the spreading inductance of deck 12.
A separate cavity for a sensitive rail keeps its modes to itself. It also creates the plane discontinuities of deck 02, so it trades one mechanism for another and has to be planned rather than improvised.
The four routes deliver their damage into budgets the earlier decks already built, and it is worth being explicit about which entry each one lands in, because the commonest error is to count a disturbance twice.
| Route | Arrives as | Enters which budget | Computed where |
|---|---|---|---|
| Through the transmitter | Amplitude modulation of the swing | The voltage noise term, and its jitter equivalent through the slew rate | Deck 10, slides 02 and 03 |
| Through the oscillator | Periodic jitter at the ripple frequency | The deterministic jitter term, bounded | Slide 03 of this deck |
| Through the return path | A series impedance at the transition | Insertion loss and reflection — the channel itself | Deck 02, deck 12 |
| Through the cavity | Crosstalk uncorrelated with the victim | Bounded uncorrelated jitter, and the crosstalk variance | Deck 07, deck 10 |
Note that only the second is a jitter term by origin. The first becomes one through the slew rate, the third is not a noise term at all but a change to the channel, and the fourth is crosstalk that happens to have arrived by an unusual route. Deck 17's compliance calculation sees all four, but it sees them in three different places.
A power-integrity analysis that stops at an impedance profile has not said whether the link works, and a signal-integrity analysis that treats the supply as ideal has not said whether the channel is the one it modelled. The two meet at precisely these four places and nowhere else, which makes the list short enough to check.
| Mechanism | Expression | Worth remembering |
|---|---|---|
| Simultaneous switching noise | $L\,N\,\Delta I / t_r$ | Linear in driver count; near zero for differential |
| Supply to jitter | $\Delta\phi = K_{\mathrm{VCO}}V_n/f_m$ | Filtered by the loop below its bandwidth |
| Worst tone frequency | At the loop bandwidth, for a type-II loop; nowhere in particular for a type-I one | Place the regulator's frequency deliberately, and state the loop order before quoting the rule |
| Reference-plane change | Cavity impedance in series | A PI property that is a SI parameter |
| DC blocking capacitors | $Z \to \omega L_{\text{mount}}/N$ | Two jobs, two different counts |
| Cavity coupling | Modal, not local | Distance does not help; damping does |
| Double counting | — | Four routes, three budget entries; count each once |
Deck 14 of eleven in Signal Integrity & High-Speed Digital Design. Every figure on this page is computed by si_models/deck14 and embedded as data; nothing is typed in by hand.
Single-page HTML · KaTeX-rendered maths · no build step. Source on GitHub