Electronics Guide

Articles in This Category

Why Channels Lose Energy

Two mechanisms dominate the loss of a printed circuit board interconnect, and they scale with frequency in different ways. Conductor loss arises from the finite resistance of the copper. At high frequency the skin effect confines current to a thin layer near the surface of the conductor, shrinking the effective cross section and raising resistance; this loss rises roughly with the square root of frequency. Surface roughness, deliberately imparted to copper foil so that it adheres to the laminate, lengthens the path that this surface current must follow and inflates conductor loss further as frequency increases. Smooth-profile foils such as reverse-treated and hyper-very-low-profile copper exist specifically to recover that penalty.

Dielectric loss arises from the laminate itself. The alternating electric field of the signal repeatedly polarizes the insulating material, and a fraction of that energy is dissipated as heat. The severity is captured by the loss tangent, or dissipation factor, of the material, and dielectric loss rises approximately in direct proportion to frequency. Because dielectric loss grows linearly while conductor loss grows only with the square root of frequency, the dielectric term overtakes the conductor term at higher frequencies and eventually dominates the channel.

This crossover is exactly why material selection matters so much in fast designs. Dissipation factor is itself frequency-dependent, so laminates must be compared at the same test frequency to be compared fairly. Standard FR-4 carries a dissipation factor near 0.02 at 1 GHz, while a low-loss digital laminate such as Panasonic Megtron 6 is specified at about 0.002 at 1 GHz and about 0.004 at 12 GHz. Dedicated high-frequency laminates go further still: Rogers RO4350B is specified at 0.0037 at 10 GHz with a dielectric constant of 3.48. Choosing a low-loss laminate can therefore cut the dielectric contribution by close to an order of magnitude, extending the reach of a link before equalization is even applied.

Several smaller contributors round out the budget. Impedance discontinuities at vias, connectors, and package escapes reflect energy rather than dissipating it, but from the receiver's point of view the effect is the same loss of usable signal, and the resulting resonances carve notches into the frequency response. Coupling into neighboring nets removes energy from the aggressor and injects noise into the victim. Radiation from unshielded structures, fiber weave effects that arise because glass and resin have different dielectric constants, and moisture absorbed by the laminate all add further, generally smaller, penalties. For a full treatment, see Channel Loss Mechanisms.

Measuring and Budgeting Loss

Loss is quantified as insertion loss, the reduction in signal amplitude between the input and the output of the channel, expressed in decibels and plotted against frequency. In S-parameter terms it is the magnitude of S21 for a single-ended channel, or of SDD21 for the differential-to-differential term of a mixed-mode measurement. A vector network analyzer measures it directly; time-domain methods derive it from a transmitted step or pulse. Because fixtures, probes, and launch structures contribute loss of their own, careful calibration and de-embedding are prerequisites for any credible number.

The single figure engineers quote most often is insertion loss at the Nyquist frequency, which equals half the symbol rate. A 32 GT/s two-level link signals at 32 gigabaud and has a Nyquist frequency of 16 GHz. Notably, a 64 GT/s four-level link also signals at 32 gigabaud, so it shares the same 16 GHz Nyquist frequency; the higher data rate is bought with additional amplitude levels rather than additional bandwidth. This is why loss budgets do not simply halve from one generation to the next. PCI Express 5.0 allows on the order of 36 decibels of bump-to-bump insertion loss at 16 GHz, whereas PCI Express 6.0 tightens the comparable budget to roughly 32 decibels, because four-level signaling demands a better signal-to-noise ratio at the slicer.

As a rough benchmark, a narrow stripline on standard FR-4 dissipates on the order of one decibel per inch at 10 GHz, while the same geometry on a low-loss laminate loses a few tenths of a decibel per inch. Exact figures depend on trace width, copper roughness, and stackup, so a real budget is built from measured or simulated data rather than a rule of thumb. The budget must then be apportioned across the whole path: the silicon package, the breakout region, the board trace, every via transition, the connector, and any cable assembly each claim a share.

A smooth loss curve is not the whole story. Standards bodies therefore define composite figures of merit that capture how well behaved a channel is, not merely how lossy it is. Insertion loss deviation measures how far the measured response departs from a smooth fitted curve, exposing the resonances that reflections create. Integrated crosstalk noise aggregates the coupling from all neighboring lanes into a single voltage. Channel operating margin, adopted in several IEEE 802.3 clauses, folds loss, reflections, crosstalk, and a reference equalizer into one decibel-valued pass or fail criterion. These metrics and the instruments behind them are developed in Channel Characterization.

From Loss to Intersymbol Interference

The practical danger of frequency-dependent loss is not the loss of amplitude alone but its effect in the time domain. A channel that attenuates high frequencies behaves as a low-pass filter, so a sharp transmitted edge arrives at the receiver stretched in time. When that stretched response of one symbol still lingers as the next symbols arrive, the energy of past bits adds to or subtracts from the present bit. This is intersymbol interference.

The clearest way to see it is the pulse response: the waveform that emerges when a single symbol-wide pulse is launched into the channel. Sampling that response at one-unit-interval intervals yields a set of cursors. The largest sample is the main cursor, the one the receiver intends to slice. Samples that follow it are post-cursors, the lingering tail that trails a transition. Samples that precede it are pre-cursors, energy that arrives ahead of the main pulse as a consequence of the channel's phase response. On a lossy channel the main cursor may hold only a fraction of the total pulse energy, with the remainder spread across a long tail of post-cursors. Equalizer tap counts are chosen directly from the length and shape of that tail.

Engineers visualize the combined effect with the eye diagram, formed by overlaying many unit intervals of the received waveform. As loss and intersymbol interference grow, the eye closes vertically, eroding voltage margin, and horizontally, eroding timing margin, until the receiver can no longer reliably distinguish one level from another. Because the interference depends on the particular sequence of bits that preceded the one being sampled, it also appears in the timing domain as data-dependent jitter, one of the deterministic components of a timing budget. The whole purpose of equalization is to reopen that eye by reshaping the channel response so that each symbol is decided cleanly, with the contributions of its neighbors removed.

Equalization Strategies

Equalization compensates loss by applying a frequency response that is, ideally, the inverse of the channel's: it boosts the high frequencies the channel attenuates, or it cancels the residual intersymbol interference directly. The compensation can be passive or active, and it can sit at the transmitter, at the receiver, or at both ends of the link.

Passive Equalization

A passive equalizer is a fixed network of resistors, capacitors, and sometimes inductors that attenuates low frequencies more than high frequencies, flattening the overall response without supplying gain. It is simple, consumes no power, adds no active noise, and is well suited to moderate loss budgets, cable assemblies, and legacy interfaces where no adaptive silicon is available. Its limitations are that it cannot add energy, so it reduces the overall signal level, and that its fixed shape cannot adapt to channels that vary from board to board, from lot to lot, or with temperature.

Feed-Forward Equalization (FFE)

Feed-forward equalization is a finite-impulse-response filter, most often placed in the transmitter, that pre-distorts the outgoing signal. By splitting the drive into weighted, time-shifted taps, the transmitter emphasizes the high-frequency edge content and de-emphasizes the steady levels, so that after the channel rolls the spectrum off, the waveform arrives with a flatter response. Because it acts on known transmitted data, transmitter FFE can address both pre-cursor and post-cursor interference, which no decision-feedback structure can do for pre-cursors. Its cost is reduced transmitted swing, since de-emphasis lowers the average launched amplitude; the transmitter spends signal-to-noise ratio to buy a flatter response. Practical transmitters use a small number of taps, commonly one pre-cursor and one or two post-cursor taps, with more available in the most demanding standards.

Continuous-Time Linear Equalization (CTLE)

A continuous-time linear equalizer is an analog filter in the receiver that applies peaking gain near the Nyquist frequency while attenuating low frequencies, again inverting the channel's low-pass character. It is typically realized as a source-degenerated differential amplifier whose degeneration network places a zero below the pole pair, producing the desired high-frequency boost. CTLE is compact and power-efficient, handles a wide span of loss, and can be adapted in coarse steps to the measured channel. Its principal drawback follows from linearity: it boosts noise, crosstalk, and reflected energy along with the wanted signal, so pushing the peaking too hard on a very lossy channel degrades the signal-to-noise ratio rather than improving it.

Decision-Feedback Equalization (DFE)

A decision-feedback equalizer is a nonlinear receiver structure that cancels interference using the receiver's own past decisions. After each symbol is sliced, its known value is scaled by tap weights and subtracted from the incoming signal, removing the trailing interference that earlier symbols impose on later ones. The defining advantage is that, because a DFE subtracts a clean, already-decided level rather than amplifying the analog input, it cancels intersymbol interference without amplifying noise or crosstalk.

Its limits follow from the same mechanism. A DFE can cancel only post-cursor interference, never pre-cursor, because future symbols have not yet been decided. An incorrect decision feeds the wrong correction into the following symbols, a failure mode known as error propagation, which becomes more troublesome as pre-correction error rates rise. Most restrictive of all, the first feedback tap must be computed and subtracted within a single unit interval, a loop that becomes impossible to close at the highest symbol rates. The standard remedy is loop unrolling, also called speculative or partial-response feedback: the receiver evaluates both possible outcomes of the first tap in parallel and uses the previous decision merely to select which precomputed result to keep, moving the critical path out of the feedback loop.

Adaptation and Link Training

Because manufacturing spread, temperature, and cable length all move the channel, modern equalizers are not set by hand but converged automatically. At link bring-up, the two ends exchange training patterns while the receiver evaluates candidate settings. PCI Express defines a set of transmitter presets that the receiver may request during link equalization, and IEEE 802.3 backplane and copper-cable clauses define a training frame that carries coefficient increment and decrement requests back to the remote transmitter. Receiver-side CTLE peaking and DFE tap weights are then adapted continuously, commonly with a least-mean-square or sign-sign least-mean-square update that drives the error between the sampled value and the decided level toward zero. This adaptive behavior is what allows one silicon design to serve short package-to-package hops and long backplane channels alike.

Building a Complete Link

No single technique covers every channel, so high-speed transceivers combine them. A typical link pairs transmitter FFE with receiver CTLE and DFE: the FFE and CTLE invert the bulk of the linear loss and open the eye enough for reliable slicing, while the DFE cleans up the residual post-cursor interference without paying the noise penalty of further linear boost. Because transmitter FFE and CTLE both flatten the channel's low-pass response, they are partly redundant, and the most demanding designs allocate gain carefully between them rather than maximizing both. Longer channels add receiver-side FFE as well, and the hardest of them turn to maximum-likelihood sequence estimation, which chooses the most probable transmitted sequence rather than deciding each symbol in isolation, at a substantial cost in power and area.

Where equalization alone cannot close the budget, retimers and redrivers extend the reach. A redriver is an analog device that equalizes and reamplifies the signal but passes jitter through; a retimer recovers the clock, makes decisions, and launches a clean signal, resetting the jitter and interference budget at the cost of latency and power. Beyond a few tens of decibels of copper loss, the pragmatic answer is often to leave copper entirely for an optical link.

The right split among all these resources is found through link-budget analysis and statistical eye simulation, which weigh loss, reflections, crosstalk, and jitter together to confirm that the equalized eye meets the standard's mask with adequate margin. Statistical methods matter here because target error ratios of one in a trillion or lower cannot be confirmed by simulating individual bits in any practical time.

Higher-Order Modulation and Error Correction

Four-level pulse amplitude modulation, universally abbreviated PAM4, carries two bits per symbol and therefore halves the symbol rate needed for a given data rate. Halving the symbol rate halves the Nyquist frequency, which on a steeply sloping loss curve can recover many decibels of insertion loss. The compensation is paid at the slicer: three stacked eyes divide the same peak-to-peak swing into thirds, and each eye is one-third the height of the equivalent two-level eye. That amounts to an intrinsic signal-to-noise penalty of about 9.5 decibels before any other impairment is considered, and it makes the receiver far more sensitive to amplitude nonlinearity, level mismatch, and crosstalk than a two-level receiver would be.

Because equalization alone cannot recover that margin, PAM4 links pair it with forward error correction. IEEE 802.3 specifies Reed-Solomon RS(544,514) coding, widely called KP4, for its high-rate PAM4 interfaces; the code appends thirty parity symbols to a 514-symbol payload and corrects up to fifteen symbol errors per codeword, taking a pre-correction error ratio in the range of one in a thousand down to one in a trillion or better after correction. PCI Express 6.0 takes a different route suited to its latency sensitivity, combining a lightweight forward error correction code with a cyclic redundancy check and link-level retry. In both cases the design target shifts: the equalizer no longer has to deliver a pristine eye, only an eye good enough for the error-correcting code to finish the job. That shift interacts directly with DFE error propagation, since bursts of correlated errors are harder for a code to absorb than isolated ones.

Conclusion

Channel loss is physics, not a defect, and it grows relentlessly with frequency. An accurate understanding of where the energy goes, a sound measurement of the channel, a well-chosen mix of passive and active equalization, and, at the highest rates, a modulation and coding scheme matched to the remaining margin together let a link operate far beyond the data rate the bare channel would allow. Getting that chain right is what separates a design that closes its budget from one that fails compliance late and expensively. The topics in this category develop each part of the chain in depth, from the physics of conductor and dielectric loss through characterization methods to the passive networks and active equalizers that keep modern multi-gigabit links open.

Related Topics