How a standard decides a channel is legal without knowing which transceiver will be plugged into it — run on the same backplane the equalisation deck could not close, and giving a different verdict for a reason worth understanding.
A standards body has to publish a test that decides whether a channel is legal. It cannot simulate the receiver that will eventually be plugged in, because that receiver is proprietary, differs between vendors, and in many cases has not been designed yet. Nor can it simply set a limit on insertion loss, because deck 16 has just shown that insertion loss does not determine the eye.
The resolution is to define a reference receiver — a specific, published, deliberately unambitious equaliser — run it against the measured scattering parameters of the channel, and report a single number. Channel operating margin is that number. A channel that works with the reference receiver is assumed to work with any real receiver, because any real receiver is at least as good.
The reference receiver is allowed to adapt. It optimises its own equaliser settings, its sampling phase and its transmit preset against each channel, because a fixed setting would fail channels that work perfectly well. So the test is not "does this channel suit our assumed receiver" but "is there any setting of a modest receiver that closes this channel", which is a much better proxy for the real question.
The published method — annex 93A of IEEE 802.3 and its descendants — is long and detailed, and it is worth stating plainly which parts are reproduced here so that the numbers on the following slides are read correctly.
A three-tap transmit equaliser chosen from a grid of presets; a continuous-time equaliser chosen from a grid of settings and normalised at Nyquist so the grid trades shape rather than gain; a decision-feedback section of limited length with limited tap magnitude; optimisation over sampling phase; and noise terms for residual intersymbol interference, crosstalk, jitter and receiver noise. Residual interference is treated statistically, because the data is random and the residual taps contribute a variance.
Explicit convolution of probability distributions rather than addition of variances; the published transmitter and package models; the full set of compliance masks; the specific numerical constants and grids of any one standard. A certification laboratory needs all of those and this is not one.
What survives the simplification is the structure of the argument, which is what this deck is about: a compliance verdict is a joint statement about a channel and a reference receiver, and reading it as a property of the channel alone is a mistake.
Run against the twenty-eight-point-eight-inch backplane channel from SerDes_Equalisation, back-drilled, the reference receiver finds the following settings and the following margin.
Here is the result that makes this deck worth writing. The equalisation deck worked the same channel by hand, with a fuller equaliser, and concluded that it falls one and a third decibels short of an error ratio of one in a million million. The channel operating margin computed here says the channel passes comfortably. Both are right, and the reason they differ is the most useful thing in this series to understand about compliance.
A margin of three decibels is a pass because the threshold is calibrated against a detector error ratio in the region of one in ten thousand to one in a hundred thousand — the error rate at the slicer, before forward error correction. Every standard that uses this method also mandates coding, and the coding carries the link from there to one in a million million and beyond. The hand budget in the equalisation deck asked for one in a million million uncoded. Two different questions, two different answers, one channel.
Deck 04 argued that an un-backdrilled stub is fatal and computed the excess loss it causes. Here is the same change expressed in the currency a standard uses.
This is what makes the channel operating margin useful as a design tool rather than only as a gate. It converts every physical decision in the preceding ten decks — the laminate, the copper foil, the antipad, the back-drill depth, the lane spacing — into a single number in a single currency, so that they can be compared against each other and against the cost of more silicon.
The channel does not change between the rows below. Only the reference receiver's assumptions change, and each variation is within the range real silicon actually spans.
The standards fit a smooth curve to the insertion loss and specify a limit on the deviation from it rather than on the loss itself. The reason is exactly the argument of deck 16. The smooth part of the loss is what the equaliser is designed to remove and a channel is allowed a great deal of it. The wiggle on top is caused by reflections, reflections put energy at particular delays, and the equaliser may not reach them.
Alongside the margin calculation, a handful of limits are applied directly to the scattering parameters. They are cheaper to check, they catch gross errors before anything is simulated, and each one corresponds to a mechanism from an earlier deck.
Slide 06 showed the number moving by several decibels under variations in the reference receiver that no channel change accompanied. A channel that passes has passed with that receiver, and interoperability arguments between vendors are very often arguments about which reference is authoritative rather than about any measurement.
The threshold corresponds to a pre-correction error ratio. A link designed without forward error correction, or with a weaker code than the standard assumes, does not inherit the verdict — as slide 04 demonstrates on this very channel.
On slide 06 the crosstalk variation moves the margin further than any of the receiver parameters. A compliance figure computed with optimistic crosstalk is optimistic by more than the margin it reports, and crosstalk is the hardest term to measure honestly because it needs every aggressor driven.
A single number cannot tell a designer which of the eleven mechanisms in this series is responsible. That is what the preceding decks are for, and it is why a failing margin should send you back to the pulse response rather than to a bigger equaliser.
Eleven decks have taken one physical question — can this signal get from that chip to this one intact — and decomposed it into mechanisms that can each be computed. It is worth collecting what each contributed to the channel they all share.
| Deck | What it establishes about this channel |
|---|---|
| 01 Transmission lines | It is many wavelengths long, in the overlap of the skin-effect and dielectric-loss regions, and well below waveguide onset |
| 02 Return paths | Its return current runs in a band a few trace heights wide, and a plane gap would cost more than the entire equalisation budget |
| 03 Materials | The laminate choice is worth tens of decibels over this length, and a constant-Dk model of it is not physically realisable |
| 04 Vias | Back-drilling is the difference between a working channel and a notch at Nyquist |
| 05 Differential signalling | Tight coupling would cost over a decibel here to buy something worth less |
| 06 Crosstalk | Its contribution cannot be equalised and is worth more than extra decision-feedback taps |
| 11–14 Power integrity | Supply ripple reaches the eye through the loop's rejection, worst at the loop bandwidth in a type-II loop |
| 07–10 Jitter | The dominant timing term is data-dependent, so it is the equaliser's problem and not the oscillator's |
| 15 Timing | The reason this is a serial link at all, rather than thirty-two slower wires |
| 16 Measurement | Whether any of the above can be believed depends on four properties of the data file |
| 17 Compliance | And the verdict depends on which question was asked |
| Quantity | Meaning | Worth remembering |
|---|---|---|
| COM | $20\log_{10}(A_s/\sigma_{\text{total}})$ | A ratio of signal to noise after a reference receiver has adapted |
| Pass threshold | 3 dB | Calibrated against a pre-correction error ratio, not against $10^{-12}$ |
| Detector error ratio | The slicer's error rate | Forward error correction takes it the rest of the way |
| ILD | Insertion loss minus a smooth fit | Isolates reflections from loss, because only the loss is equalisable |
| Return loss mask | $|S_{11}|$ limit over the band | Bounds the reflections that produce the ILD |
| ICR / ICN | Crosstalk against insertion loss | The term that moves the verdict most |
| The caveat | — | A verdict is a joint statement about a channel and a reference receiver |
Deck 17 of eleven in Signal Integrity & High-Speed Digital Design. Every figure on this page is computed by si_models/deck17 and embedded as data; nothing is typed in by hand.
Single-page HTML · KaTeX-rendered maths · no build step. Source on GitHub