Reporting Signal and Power Integrity Margin

In hardware design reviews, few slides receive more attention than the final sign-off table: “PCIe Gen 5 Lane 0: +14 ps timing margin, +22 mV voltage margin — PASS.” To project managers and layout teams, those green numbers are treated as an objective, binary warranty that the board will function reliably in the field.

To an experienced signal integrity engineer, however, an isolated margin number is completely meaningless. Was that 14 picoseconds evaluated at typical room temperature silicon, or at slow-weak process corners with a 105 °C junction temperature and 5% supply droop? Did the simulation include the sixteen adjacent data lines switching simultaneously, or was the channel simulated in total isolation? Did the statistical model extrapolate to a Bit Error Rate of 10^-6 or 10^-12?

Margin is not a static physical property of a circuit board. It is the mathematical distance between a simulated waveform and an interface specification, conditioned on an explicit stack of assumptions. A margin report that fails to state its operating corners, statistical stacking conventions, and conscious model omissions does not communicate confidence — it creates a dangerous illusion of safety that unravels the moment hardware reaches high-volume production.

What a margin is a statement about

Margin is the distance between what a design achieves and what it must achieve. Both halves of that are less definite than they sound.

"What it achieves" depends on which corner you evaluated, how you combined the variations, and which impairments were in the model. "What it must achieve" depends on the target error rate, the reference plane and the reference receiver.

So a margin of 5 ps is not a fact about a board. It is a fact about a board plus a set of choices, and the choices move the answer more than most design changes do. Reporting the number without them is the problem this page is about.

Corner stacking is a choice, and it must be stated

There is no neutral way to combine variations. The two common conventions differ by a large factor and answer different questions:

Worst-case stacking. Every variation at its bad extreme simultaneously. Guaranteed conservative, and for a channel with many independent contributors it describes a part that will essentially never be built. Useful as a bound; misleading as a prediction.
Root-sum-square. Combines statistical quantities — standard deviations of distributions — into a number that reflects a realistic population. It needs two conditions that are routinely broken. It needs actual distributions: a min/max specification limit is not a sigma, and RSS-ing datasheet limits yields a figure with no defined confidence attached to it. And it needs independence: when two terms share a cause — a fast driver and a fast edge both come from the same process corner — the covariance term is non-zero and RSS understates the spread. State the distributions assumed, the confidence claimed, and which terms were treated as independent so another engineer can check the result.

Neither is right in general. What matters is that the report says which was used, for which contributors, and why — because a reader who knows the convention can adjust, and a reader who does not cannot.

The single most useful line in a margin report What was not included. No crosstalk. Nominal package model. Typical PDN noise. Two aggressors rather than sixteen. Every simulation excludes something, and naming the exclusions is what lets the next person judge whether the number applies to their question. A report with no exclusions listed has not been thought about — it has not become complete.

What makes a number actionable

A margin figure should come with enough context to act on it without asking a question:

Where it was measured — which node, before or after equalisation, at what reference plane.
At what BER, since eye height and width mean nothing without it, and whether that number was measured or extrapolated.
Which corners, and how they were combined.
What is in the model: which blocks are extracted, which are estimated, which are vendor-supplied and unvalidated.
The sensitivity. A number alone is a verdict; a number with its sensitivity is a design tool. “3 ps of margin, losing 0.4 ps per additional aggressor” tells the layout engineer what to do.

One cultural note that matters more than any of the above. A report that always says “pass” stops being read. The value of the function is that it occasionally says the uncomfortable thing early, with the evidence attached — and that only survives if a marginal result is reported as marginal rather than rounded into compliance. Reporting a negative margin with the sensitivity attached is more useful to a programme than reporting a positive one that assumed the problem away.

Why this decides how your board behaves

Incomparable numbers cause wrong decisions. A worst-case margin and an RSS margin differ by a large factor, and a team comparing them is comparing nothing.
A number without its assumptions cannot be re-used. A later design variant needs a durable record of the assumptions behind the result.
Hidden pessimism costs money. Stacking every variation at its limit describes a board that will never be built, and designing to it buys layers and laminate you did not need.
Hidden optimism costs more. Combining deterministic quantities in quadrature understates the total, and the failure appears in the field.

What to do about it

State the stacking convention, always. Worst-case or RSS, and which contributions went into which. This single line prevents most of the confusion.

Never combine deterministic and random in quadrature. RSS is for independent statistical quantities. Deterministic contributions add directly, worst case to worst case — and two random contributions driven by the same supply noise are not independent either.

Report the model version and the inputs. Which S-parameter files, which buffer models, which corners, which data rate. A margin is reproducible or it is an opinion.

Say what is not in the number. Every analysis omits something — supply noise, an unmodelled connector, temperature. Naming the omissions is what lets a reviewer judge the margin rather than accept it.

What makes a number actionable A margin report that someone can act on says, at minimum: the quantity and its units; the reference plane; the target error rate; the corner set and stacking convention; the model versions; and the known omissions. Six lines. A number without them is not a result someone else can build on, however carefully it was computed.

Go deeper — why the two conventions differ so much, and what to do instead

Worst-case stacking puts every variation at its bad extreme simultaneously. It is guaranteed conservative and, for a channel with many independent contributors, describes a part that will essentially never be built. Useful as a bound; misleading as a prediction.

Root-sum-square combines standard deviations of independent distributions:

σ_total = √( σ₁² + σ₂² + … ) valid only for independent random quantities — not for deterministic ones, and not for correlated ones

The gap between the two conventions grows with the number of contributors. With four equal independent terms, worst-case is 4× one of them and RSS is 2×. With nine, it is 9× against 3×. That is why the convention has to be stated: the ratio is not a detail.

What to do instead of choosing. Report both where it matters, or run a statistical analysis that samples the actual distributions with their correlations. The latter is the honest answer when the decision is expensive, because it neither invents a board that cannot exist nor assumes independence that is not there.

The correlations that bite. Several variations on this site are not independent, and treating them as such is a common quiet error: resin flow and D_k are both driven by the glass-to-resin ratio; several jitter contributions share one supply; temperature moves loss, delay and capacitance together. Each of those makes RSS optimistic.

In the real world

Margin reporting is where technical work becomes an organisational artefact. The number leaves your hands and gets compared, aggregated and decided upon by people who were not there — and it will be read as more definite than it is.

Which makes the assumptions block the most valuable part of the report, and the part most often cut for brevity. Six lines of context is what turns a number into evidence, and evidence is what survives being questioned six months later.