The Channel Fights Back: Loss and Intersymbol Interference

The copper between two friends

Picture two people in a long stone corridor. One whispers a crisp, clipped word; by the time it reaches the other end, the hard consonants have softened into a blurry hum and the echoes of the previous word are still ringing as the next one arrives. That corridor is exactly what a SerDes channel does to your bits. The serializer at one end launches a beautifully square voltage edge, but the wire delivers something rounded, delayed and overlapping at the other end.

The "channel" is not abstract. It is a physical chain: a ball or bump leaving the transmitter package, a breakout trace on the printed-circuit board, a via diving between PCB layers, a connector, maybe a cable, another connector, more trace, another via, and finally the receiver package. Every one of those is copper surrounded by insulating dielectric, and every one of them taxes the signal. A modern 56 Gb/s PAM4 link can lose more energy in 30 cm of FR-4 board than a radio link loses crossing a room.

Why a wire becomes a low-pass filter

Two distinct physical effects gang up on high frequencies, and it pays to keep them separate in your head because they scale differently.

Skin effect is the first. At DC, current spreads evenly across the whole cross-section of the copper. As frequency rises, the magnetic field inside the conductor pushes current toward the outer surface — the "skin" — leaving the core idle. The effective conducting area shrinks, so the effective resistance climbs. Because the skin depth shrinks as 1/√f, this loss grows roughly with √f. The same physics you may have met in skin effect on power busbars is here squeezing your data edges.

Dielectric loss is the second. The insulating material around the copper — FR-4 glass-epoxy, or a lower-loss laminate like Megtron — is full of polar molecules. The signal's changing electric field tries to wiggle them, and they rub against their neighbours, turning a sliver of your signal into heat on every cycle. This loss grows roughly linearly with f, so it dominates at the very top of the band. A material's badness here is captured by its loss tangent, tan δ; halving tan δ is one of the most direct ways to buy back loss.

  Attenuation vs frequency  (the two enemies add up)

  loss
  (dB) ^
       |                                     total ___
       |                                   ___/
       |                            ___/···  dielectric  (∝ f)
       |                      ___/   ···
       |                ___/    ·········   skin (∝ √f)
       |           __/  ··········
       |       __/·······
       |    _/····
       |  _/··
       |_/_____________________________________________> frequency
       0                         f_Nyquist

  Below f_Nyquist the channel is gentle; right where your data
  lives, the curve is already plunging.

Insertion loss climbs monotonically with frequency: √f from skin effect plus linear-in-f from the dielectric. The fast edges that define your bits are made of exactly the high frequencies the channel punishes most.

When one bit bleeds into the next

Now watch what that rounding does in time. Send a single, isolated '1' pulse one bit-period wide into a lossy channel. Out the far end it does not come back as a clean rectangle — it comes back lower, wider, and dragging a long tail. The energy the channel could not deliver on time arrives late, spilling into the time-slots of the bits that follow. This spreading is the root cause; the symptom it produces has a name.

  IDEAL (no loss)            LOSSY CHANNEL (pulse response)

  v |  ___                    v |    _
    | |   |                     |   / \__               main cursor
    | |   |                     |  /     \___           + long tail
    |_|   |____                 |_/          \_____      bleeds into
    +----+----+--> t            +----+----+----+--> t   next slots
     bit  next                   bit  next  next

  cursors sampled at bit centers:
     ideal : [ ... 0   1   0   0 ... ]
     lossy : [ ... .1  .6  .25 .1 ... ]   <- neighbors are nonzero!
            (pre-cursor)↑  main↑  (post-cursors)↑

A single pulse spreads in time. Sampled at the centre of each bit, the lossy response is nonzero in the neighbouring slots — those are the ISI 'cursors'. The big one is the main cursor; the leftovers before and after are pre- and post-cursors.

That is exactly intersymbol interference (ISI): the voltage the receiver sees at the sampling instant for bit *n* is not just bit *n*, but a sum of bit *n* plus leaked tails from earlier and later bits. Whether the receiver's threshold comparator reads a '1' or a '0' now depends on what the neighbouring bits were. The link has developed a memory — and a bad one, because that memory corrupts the very decision it is feeding.

How ISI closes the eye

The cleanest way to *see* ISI is the eye diagram from rung 1. Overlay thousands of received bit-periods on top of each other, all aligned to the clock, and the traces carve out an open 'eye'. The eye's height is your voltage margin — how far above the noise the '1' and '0' levels separate. Its width is your timing margin — how much the sampling clock can wander and still land in the clear.

ISI attacks both dimensions at once. Because each bit's level now depends on its neighbours, the '1' rail is no longer one clean line but a fuzzy band of slightly-different '1' levels — and likewise for '0'. The band eats into the eye's centre from top and bottom, shrinking the height. Meanwhile the spread tails make some edges cross the threshold early and others late, so the crossings smear sideways and the eye narrows in width. Push the loss far enough and the eye slams shut: there is no instant and no voltage at which a simple comparator can reliably tell '1' from '0'.

  LOW LOSS  (eye wide open)        HIGH LOSS  (ISI closing it)

   1 \__________/  \______          1 \____    ____/\___
         \    /    \    /                 \\  //   \\
          \  /      \  /                   \\//     ====   <- eye
   -----    \/        \/   <-thresh  -----  XX      ====      barely
          /  \      /  \                   //\\     ====      open
         /    \    /    \                 //  \\   //
   0 ___/      \__/      \__       0 ____//    \\__\\___
      <-- 1 UI -->                    height ↓   width ↓

  Sampling point wants the WIDEST, TALLEST opening it can find.

Same data, two channels. Loss-driven ISI shrinks the eye in both height (voltage margin) and width (timing margin). A closed eye means a raw bit-error rate too high to use — before any equalization is applied.

Budgeting loss in decibels at Nyquist

Engineers do not argue about channels in vague terms; they put a number on the damage. That number is insertion loss, measured in decibels and quoted at the Nyquist frequency — the fundamental frequency of the fastest pattern, a repeating 1010… stream. For non-return-to-zero (NRZ) signalling, that pattern repeats every two bits, so f_Nyquist is exactly half the bit rate (the symbol rate, the baud, divided by two).

  Worked example: a 32 Gb/s NRZ link

    bit rate          = 32 Gb/s
    f_Nyquist         = 32 / 2 = 16 GHz      <- spec the loss HERE

    measured channel insertion loss @ 16 GHz = -24 dB
      -> received fundamental is 24 dB down
      -> 10^(-24/20) ≈ 0.063  -> only ~6.3% of the amplitude survives!

  Reading a spec line:
    "channel budget: ≤ 30 dB at Nyquist (10 GHz / 5 GHz / etc.)"
    means: equalization + good material must keep total loss
    under 30 dB at the half-baud frequency, or the eye won't open.

  PAM4 twist: 4 levels squeeze 2 bits/symbol, so f_Nyquist is
  HALF that of NRZ at the same bit rate — but the eye is split into
  three thinner sub-eyes, so PAM4 is far less loss-tolerant per dB.

Insertion loss is specified in dB at the Nyquist frequency because that is where the channel hurts the fundamental of your worst-case toggling pattern. 24 dB sounds modest until you convert it: only ~6% of the swing arrives.

This single number lets a system architect partition responsibility before any board is built. The 'loss budget' says, for example, that the package may eat 4 dB, the PCB traces 18 dB, two connectors 6 dB, and so on — and that the receiver's equalizer must be strong enough to claw back the total. As interconnect scaling keeps shrinking on-chip wires while board distances stay stubbornly long, more and more of the link's difficulty lives in this off-chip channel, which is why so much SerDes innovation targets it.

Find the symbol (baud) rate and the modulation (NRZ vs PAM4) — this fixes the Nyquist frequency.
Read or measure the channel's insertion loss in dB at that Nyquist frequency, not at DC.
Compare it against the SerDes IP's published loss budget — the maximum dB it can equalize and still close the link.
If you are over budget, attack the channel: lower-loss laminate (smaller tan δ), shorter or wider traces, fewer vias, better connectors — before leaning harder on equalization.

The enemy that equalization exists to defeat

Step back and the strategy of the whole track snaps into focus. The channel is a fixed, frequency-shaping low-pass filter that you usually cannot change much once the box is built. ISI is the time-domain shadow it casts. The closed eye is the verdict. And insertion loss at Nyquist is the scoreboard. Everything that comes next is a counter-move against this one opponent.

If the channel is a low-pass filter, the obvious cure is to build a high-pass-ish filter that does the opposite — boosting exactly the high frequencies the channel starved. That is what equalization is, and it comes in several flavours you will meet in rung 3: a transmit filter that pre-distorts the signal so it arrives correct, and receive filters that either lift the high band continuously or actively subtract the leaked tails bit by bit. None of them creates information for free; they reshape, and they cost power and noise. But they can pry a slammed-shut eye back open.