Channel Loss and Equalization: FFE, CTLE and DFE
Show why a lossy channel closes the eye, and open it again three ways.
You Will Be Able To
- Explain how channel loss turns into intersymbol interference.
- Predict what transmit de-emphasis does to the swing of repeated bits.
- Explain what a CTLE boosts and what a DFE subtracts.
- Choose an equalizer for pre-cursor or post-cursor interference.
- Lesson
- P10.9
- Time
- 35 min
- Runs on
- ngspice
- Check
- 6 questions
A 10 Gb/s link works on a short board. On a longer one the same signal arrives with its eye less than half as tall. Nothing is broken: the channel is just lossy.
| Long bench channel | Eye height (of 1000 mV) | Eye width | What it costs |
|---|---|---|---|
| No equalizer | 466 mV | 0.82 UI | |
| Transmit FFE, c1 = 0.25 | 500 mV | 0.92 UI | Lower swing on repeated bits |
| Passive CTLE and a gain of 4 | 1004 mV | 0.90 UI | The gain stage also amplifies noise |
| 1-tap DFE, h1 = 0.12 | 638 mV | 0.90 UI | A wrong decision is fed back |
ngspice runs: 10 Gb/s PRBS7 through a 6-section RC ladder. The short channel gives 992 mV and 1.00 UI. Every part is a bench value.
Loss Becomes Interference
A board trace or a cable loses more at high frequencies than at low ones. A fast edge holds high frequencies, so it comes out slow.
- channel
- channel + CTLE + gain 4
Read the plot as a table
| Frequency | channel | channel + CTLE + gain 4 |
|---|---|---|
| 10 MHz | -0.00007663 dB | 0.000285 dB |
| 100 MHz | -0.007657 dB | 0.02834 dB |
| 1 GHz | -0.7092 dB | 1.817 dB |
| 10 GHz | -14.53 dB | -3.164 dB |
ngspice AC run. At 5 GHz, the Nyquist frequency of 10 Gb/s, the channel loses 7.8 dB. With the CTLE the gain there is +2.1 dB.
- Neutral
- Channel / handshake
- Data
- Structure
- Send one bit
The transmitter sends a sharp step one UI long.
- Lose the high frequencies
The channel passes slow changes and weakens fast ones. The edges come out slow and rounded.
- Main cursor
At the sampling instant the pulse holds most of its height. That part is the main cursor.
- Tails
The rest spills into the bits after it (post-cursor) and, a little, into the bit before (pre-cursor).
- Intersymbol interference
Each sample is its own bit plus the tails of its neighbours. A long run of ones then a single zero suffers most.
Read the diagram as text
- Sharp bit: one UI wide
- Lossy channel: low-pass
- Smeared pulse: main cursor plus tails
- Pre-cursor: leaks into the bit before
- Post-cursor: leaks into the bits after
- Sharp bit → Lossy channel
- Lossy channel → Smeared pulse
- Smeared pulse → Pre-cursor: early part
- Smeared pulse → Post-cursor: late part
TI SLYY057B p. 7 describes the lossy board the same way: it acts like a low-pass filter and shrinks the eye.
Read the eye as a table
| Eye | Height | Width |
|---|---|---|
| Eye -500 mV to 500 mV, 1 V apart | 466.4 mV | 0.82 UI |
Best sampling phase: 1.16 UI (116 ps) after each symbol is launched (the dashed line). Measured over 122 symbols after a warm-up of 10 UI. The height is the gap between the received levels at that phase; the width counts the phases, in steps of 0.02 UI, where the eye stays open.
ngspice run, folded and measured at build time.
This is intersymbol interference (ISI). The eye shrank to 466 mV and 0.82 UI. The thick bands are bits that followed different neighbours.
Cursors: The Pulse Response in Numbers
Send one bit alone and record what arrives: the pulse response. Sampled a UI apart, it gives the cursors every equalizer works on.
Read the plot as a table
| Time | channel output |
|---|---|
| 0 s | 0 V |
| 200 ps | 679.6 mV |
| 400 ps | 54.81 mV |
| 600 ps | 2.628 mV |
| 800 ps | 126 µV |
| 1 ns | 6.042 µV |
| 1.2 ns | 289.7 nV |
| 1.4 ns | 13.89 nV |
ngspice run. h0 is the peak, h1 to h3 are the tails, and h−1 is the small pre-cursor.
| Cursor | h−1 | h0 | h1 | h2 | h3 |
|---|---|---|---|---|---|
| Value (V, for a 1 V bit) | 0.019 | 0.733 | 0.193 | 0.042 | 0.009 |
Read from the pinned ngspice run at the peak and at whole UIs from it.
Any received sample is its own bit times h0 plus its neighbours times their cursors. In the worst pattern every neighbour works against it. The eye is then h0 minus the other cursors: 466 mV, the measured eye.
A transmit tap reshapes the cursors. With c1 = h1 / (h0 + h1) the first post-cursor becomes zero: c1 = 0.21 for this channel. That is the zero-forcing setting, close to the 0.18 that P10.9.1 finds best.
Transmit FFE: Shape the Bits Before They Leave
A feed-forward equalizer (FFE) at the transmitter adds a weighted copy of the neighbouring bits. The weights are called taps.
The common form is de-emphasis. A bit that repeats the one before it is sent smaller than a bit that follows a transition.
Read the eye as a table
| Eye | Height | Width |
|---|---|---|
| Eye -500 mV to 500 mV, 1 V apart | 499.9 mV | 0.92 UI |
Best sampling phase: 1.06 UI (106 ps) after each symbol is launched (the dashed line). Measured over 122 symbols after a warm-up of 10 UI. The height is the gap between the received levels at that phase; the width counts the phases, in steps of 0.02 UI, where the eye stays open.
ngspice run.
The eye grew from 466 to 500 mV and widened from 0.82 to 0.92 UI. De-emphasis lowers the low-frequency swing to match what the channel lets through at high frequency.
USB 3.2 Gen 2 uses a 3-tap FIR at the transmitter. It has a pre-cursor tap (pre-shoot), the main tap and a post-cursor tap (de-emphasis). It allows 1.2 to 3.2 dB of pre-shoot and 2.1 to 7.0 dB of de-emphasis (§6.7.5.2, Table 6-21).
TI's transmitter has the same three taps (SPRABC1 §8.3.6), and PIPE carries them as C-1, C0 and C+1 (§7.1.13 to §7.1.15).
The Coefficient Space and Presets
Standards fix a set of tap settings, presets, and let the two ends pick and fine-tune one.
USB4 lists 16 presets with their coefficients C−1, C0 and C1 (§3.1.3.1.1.2, Table 3-4). They are scaled so that |C−1| + C0 + |C1| is the full swing.
| USB4 preset | C−1 | C0 | C1 | Worst-case eye on the bench channel |
|---|---|---|---|---|
| 0 | 0 | 1 | 0 | 466 mV |
| 2 | 0 | 0.83 | -0.17 | 551 mV |
| 3 | 0 | 0.78 | -0.22 | 530 mV |
| 4 | 0 | 0.69 | -0.31 | 353 mV |
| 7 | -0.05 | 0.79 | -0.16 | 508 mV |
| 12 | -0.09 | 0.77 | -0.14 | 443 mV |
| 14 | -0.17 | 0.83 | 0 | 268 mV |
The coefficients are USB4's. The eyes are the peak-distortion bound from the pinned cursors: a lower bound, since PRBS7 rarely forms the worst pattern.
On this channel preset 2, de-emphasis with no pre-shoot, gives the best bound: its pre-cursor is tiny, so pre-shoot only costs swing.
PIPE carries the PHY's FS and LF limits as LocalFS and LocalLF (§7.2.19, §7.2.20). PCIe's rules for its coefficient space are not in our library.
CTLE: A Zero That Boosts High Frequencies
A continuous-time linear equalizer (CTLE) is a filter at the receiver. It has a zero that lifts the high frequencies, against the channel's tilt.
The bench CTLE is passive: a resistor with a capacitor across it, then a resistor to ground. At low frequency it divides by 4. At high frequency the capacitor shorts the first resistor, so nothing is lost there.
Read the eye as a table
| Eye | Height | Width |
|---|---|---|
| Eye -500 mV to 500 mV, 1 V apart | 1.004 V | 0.90 UI |
Best sampling phase: 0.68 UI (68 ps) after each symbol is launched (the dashed line). Measured over 122 symbols after a warm-up of 10 UI. The height is the gap between the received levels at that phase; the width counts the phases, in steps of 0.02 UI, where the eye stays open.
ngspice run.
With a gain stage of 4 after it, the eye is 1004 mV tall and 0.90 UI wide. Most of that height is the gain stage. The CTLE's job is the shape: flat from DC to the Nyquist frequency.
A CTLE boosts high-frequency noise and crosstalk too. Too much boost overshoots and closes the eye again.
USB 3.2 gives a reference CTLE as a transfer function with one zero and two poles (§6.8.2.1). It notes that on long channels the eye is closed without receiver equalization (§6.8.2).
TI's receiver adapts its equalizer by itself. It watches the data patterns and transition times and raises or lowers the low-frequency gain (SPRABC1 §8.2.3).
DFE: Subtract the Last Decision
A decision feedback equalizer (DFE) knows the bit it just decided. It knows how much of that bit's tail lands on the current sample, so it subtracts it.
Read the eye as a table
| Eye | Height | Width |
|---|---|---|
| Eye -500 mV to 500 mV, 1 V apart | 638 mV | 0.90 UI |
Best sampling phase: 1.10 UI (110 ps) after each symbol is launched (the dashed line). Measured over 122 symbols after a warm-up of 10 UI. The height is the gap between the received levels at that phase; the width counts the phases, in steps of 0.02 UI, where the eye stays open.
ngspice run.
One tap took the eye from 466 to 638 mV and from 0.82 to 0.90 UI. The DFE adds no noise: it subtracts a clean, decided value.
A DFE can only cancel post-cursor ISI, the tails of bits already decided. The pre-cursor belongs to a bit not yet decided, so the transmitter's pre-shoot tap handles it.
A wrong decision feeds back the wrong correction, which can cause the next error. That is error propagation.
USB 3.2 uses a one-tap reference DFE for transmitter compliance, with its coefficient between 0 and 50 mV (§6.8.2.2.2). AMD's wizard picks DFE mode above 14 dB of loss at the Nyquist frequency, and a low-power mode below (PG182).
Adaptation and DFE Timing
The best taps depend on the channel, which changes with every board and cable. Receivers therefore adapt them while data flows, as TI's receiver does (SPRABC1 §8.2.3).
Sign-sign LMS is a common rule. After each decision it nudges the tap a small step, using only the signs of the error and of the last decision.
| Symbols | DFE tap | Wrong decisions in the stretch |
|---|---|---|
| 250 | 0.012 | 10 |
| 500 | 0.090 | 3 |
| 1000 | 0.168 | 1 |
| 2000 | 0.184 | 0 |
| 5000 | 0.219 | 0 |
| 20000 | 0.199 | 0 |
Bench run on the channel's cursors with uniform noise: without a DFE, 506 of 20,000 decisions were wrong. The tap starts at 0 and settles near h1 = 0.193.
A DFE has a hard timing problem. Within one UI, 100 ps at 10 Gb/s, the last decision must be made, scaled and subtracted.
A speculative (loop-unrolled) DFE slices both candidates, y − h1 and y + h1, in advance, and the last decision only picks one. On 20,000 bench symbols it made the same decisions as the direct form: 0 differences.
Try It
The netlist runs an AC sweep of the bench channel, then of the channel with the CTLE and its gain of 4. The meas lines read both at 10 MHz and at 5 GHz.
The buffer Eb scales the channel so that it passes 0 dB at low frequency. Only the high-frequency loss is left.
Change C1 ch rx0 0.5p to 1p and run. The zero moves down, so the boost at 5 GHz grows. Then try 0.25p and watch the boost fall below what the channel loses.
Run the netlist to draw its plots here.
ch_dc ≈ -0.00007662613 ch_nyq ≈ -7.772304 eq_dc ≈ 0.0002850146 eq_nyq ≈ 2.101951Check Yourself
Common Mistakes
Wrong Say that equalization adds signal power.
An FFE lowers the swing of repeated bits, and a passive CTLE attenuates the low frequencies; neither adds energy.
Fix Equalization makes the response flat across the band. Any gain comes from an amplifier, which also amplifies noise.
Wrong Expect a DFE to cancel pre-cursor ISI.
The pre-cursor comes from a bit the receiver has not decided yet, so there is nothing to feed back.
Fix Cancel pre-cursor ISI with a pre-shoot tap at the transmitter or with a linear equalizer.
Wrong Turn the CTLE boost up until the high frequencies are tallest.
Too much boost overshoots the edges and amplifies high-frequency noise and crosstalk, which closes the eye again.
Fix Aim for a flat channel up to the Nyquist frequency, and check the eye, not only the gain.
Wrong Expect de-emphasis to raise the signal level at the receiver.
It lowers the swing of repeated bits; the eye opens because repeated bits and transition bits arrive at more equal levels.
Fix Rate it by the eye. On the bench the height rose only from 466 to 500 mV; the width grew from 0.82 to 0.92 UI.
Remember
- Channel loss rises with frequency. It smears each bit into its neighbours: intersymbol interference.
- A transmit FFE shapes the bits before they leave. De-emphasis sends repeated bits smaller.
- A CTLE boosts high frequencies with a zero. It boosts noise and crosstalk too.
- A DFE subtracts the tail of the last decided bit. It adds no noise but cannot touch the pre-cursor.
- Rate an equalizer by the eye it opens, height and width, not by its gain.