spread spectrum clock generation (SSCG) — an advanced clock generation solution for electromagnetic interference (EMI)

image-20250913195157852

image-20250913192850807

image-20250913192820120

SSC modulation profile

[https://www.synopsys.com/blogs/chip-design/understanding-pcie-spread-spectrum-clocking.html]

The most common modulation techniques are down-spread and center-spread:

  • Down-spread: Carrier is modulated to lower than nominal frequency by specified percentage, and not higher

  • Center-spread: Carrier is modulated both higher and lower than nominal frequency by specified percentage

image-20250913194151764


Steve Glaser, NVidia Corporation. Clocking Mode Terminology-2018-03-01. Workgroup: PCI Express - Protocol

Sakthivel Ramaiah, Design Engineering Architect Cadence Design Systems, PCI-SIG DevCon India 2026: Clock Modelling for Formal Verification of PCIe® Elastic Buffer

image-20251011213209060

image-20261117002944716


Mike Hertz. Configuring the SSCTrack operator for Spread Spectrum Clock Demodulation of Serial Data [https://cdn.teledynelecroy.com/files/appnotes/ssctrack_appnote.pdf]

image-20260110130021422

SSCG Architectures

K. -H. Cheng, C. -L. Hung, C. -H. Chang, Y. -L. Lo, W. -B. Yang and J. -W. Miaw, "A Spread-Spectrum Clock Generator Using Fractional-N PLL Controlled Delta-Sigma Modulator for Serial-ATA III," 2008 11th IEEE Workshop on Design and Diagnostics of Electronic Circuits and Systems, Bratislava, Slovakia, 2008 [https://sci-hub.se/10.1109/DDECS.2008.4538758]

image-20250903230209570


Due to \(f= K_{vco}V_{ctrl}\), its derivate to \(t\) is \[ \frac{\mathrm{d}f}{\mathrm{d}t} = K_{vco}\frac{\mathrm{d}V_{ctrl}}{\mathrm{d}t} \]

For chargepump PLL, \(dV_{ctrl} = \frac{\phi_e I_{cp}}{2\pi C}dt\), that is \[ \frac{\mathrm{d}f}{\mathrm{d}t} = K_{vco} \frac{\phi_e I_{cp}}{2\pi C} \]

Refclk Clocking Architectures

PCI Express Base Specification Revision 3.0

Jeff Morriss, Intel Gerry Talbot, AMD. PCI-SIG Devcon 2006. Jitter Budgeting for Clock Architecture

Verification of SRIS/SRNS Clocking [https://www.esaindia.com/emailer/download/sria-srns-white-paper-final-v3.pdf]

Common Reference Clock (CC)

Common Refclk Rx architectures are characterized by the Tx and Rx sharing the same Refclk source

Most of the SSC jitter sourced by the Refclk is propagated equally through Tx and Rx PLLs, and so intrinsically tracks LF jitter

The amount of jitter appearing at the CDR is then defined by the difference function between the Tx and Rx PLLs multiplied by the CDR highpass characteristic

image-20250719172803394 \[ H(s)= H_1(s)e^{-sT} - \left[H_1(s)e^{-sT}(1-H_3(s)) + H_2(s)H_3(s) \right] = [H_1(s)e^{-sT} -H_2(s)]H_3(s) \] where \(H_3(s)\) is similar to \(NTF_{VCO}\), \(1-H_3(s)\) is similar to \(NTF_{REF}\)

image-20250719181032685


image-20250814011504270

Data Clocked Refclk Rx Architecture

A data clocked Rx architecture is characterized by requiring the receiver's CDR to track the entirety of the low frequency jitter, including SSC

image-20250719183101724

Separate Reference Clocks with SSC (SRIS)

TITLE: Separate Refclk Independent SSC Architecture (SRIS) DATE: Updated 10 January 2013 AFFECTED DOCUMENT: PCI Express Base Spec. Rev. 3.0 SPONSOR: Intel, HP, AMD

image-20250719183242222 \[\begin{align} X_{LATCH}(s) &= X_1(s)H_1(s) - \left[X_1(s)H_1(s)(1-H_3(s)) + X_2(s)H_2(s)H_3(s) \right] \\ & = \left[X_1(s)H_1(s) -X_2(s)H_2(s)\right]H_3(s) \end{align}\] where \(H_3(s)\) is similar to \(NTF_{VCO}\), \(1-H_3(s)\) is similar to \(NTF_{REF}\)

image-20250719182447193

image-20250719182517511

image-20250719182123550

image-20250719181221513


image-20250719152821655


image-20250814011411755

Separate Reference Clocks with No SSC (SRNS)

image-20250814011354803

Refclk Jitter Filter Analysis

Greg Richmond Skyworks, Inc. July 27, 2023, 128GTps (Gen7) Refclk Jitter Filter Analysis

image-20260605231416117

image-20260605231534909

image-20260605231554054

SSC on digital CDR

Gerry Talbot, "Impact of SSC on CDR" , April 12th, 2012, PCI Express EWG

image-20250927094630403

ssc-cdr.drawio

n 0 1 2 3 ...
\(A_0\) \(0\) \(k_0\cdot \Delta\) \(k_0\cdot 2\Delta\) \(k_0\cdot 3\Delta\)
\(A_1\) \(0\) \(k_1k_0\cdot \Delta\) \(k_1k_0\cdot 3\Delta\) \(k_1k_0\cdot 6\Delta\)

\[\begin{align} f[n] &= \frac{A_1[n]-A_1[n-1]}{T} = \frac{k_1\cdot A_0[n]}{T} \\ \Delta f [n] & = \frac{f[n] -f[n-1]}{T} = \frac{k_0k_1\cdot \Delta}{T^2} \end{align}\]

the polyfit of \(f[n]\) is consistent with \(\Delta f [n]\)

image-20250927172529324

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
import matplotlib.pyplot as plt
import numpy as np


f0 = 1
fa = f0/16
fs = 32*f0

Mc = 1000
t = np.arange(0, 1/f0*Mc, 1/fs)
phi_raw = 2*np.pi*f0*t
y = np.sin(phi_raw)

plt.figure(figsize=(24,12))
plt.subplot(3, 1, 1)
plt.plot(t, y); plt.title(r'$\sin(\omega_0 t)$', fontsize='xx-large')

k0 = 0.25
k1 = 0.36

A0 = [0]
A1 = [0]
N = int(Mc*fa/f0)
delta = 0.1
dlsb = delta * 2*np.pi

f_deriv_formula = k0*k1*delta*fa**2
print(f'freq derivation by formula = {f_deriv_formula:.3e}')

for i in range(N):
A1.append(A1[-1] + k1*A0[-1])
A0.append(A0[-1] + k0*dlsb)


phi_A1 = np.ones((len(A1), int(fs/fa)))
A1_arr = np.array(A1).reshape((len(A1), 1))
phi_A1 = phi_A1 * A1_arr
phi_A1 = phi_A1.flatten()

t_phi = np.arange(phi_A1.shape[0])*1/fs

plt.subplot(3, 1, 2)
plt.plot(t_phi, phi_A1); plt.title(r'$\Delta \Phi(t)$', fontsize='xx-large')

Ntot = min(t.shape[0], t_phi.shape[0])
t_tot = t_phi[:Ntot]
phi_tot = phi_raw[:Ntot] + phi_A1[:Ntot]
y_tot = np.sin(phi_tot)

phi_tot_deriv1 = (phi_tot[1:] - phi_tot[:-1])*fs/2/np.pi
phi_tot_deriv2 = (phi_tot_deriv1[1:] - phi_tot_deriv1[:-1])*fs

f_deriv_polyfit, _ = np.polyfit(np.arange(phi_tot_deriv1.shape[0])*1/fs, phi_tot_deriv1, 1) # array([3.51305094e-05, 9.99455375e-01])
print(f'freq derivation by polyfit = {f_deriv_polyfit:.3e}')

plt.subplot(3, 1, 3)
plt.plot(t_tot, y_tot); plt.xlabel(r'$t$', fontsize=24); plt.title(r'$\sin(\omega_0 t+\Delta \Phi(t))$', fontsize='xx-large')
plt.show()

y_raw_export = np.zeros((t.shape[0], 2))
y_raw_export[:,0] = t
y_raw_export[:,1] = y

y_tot_export = np.zeros((t.shape[0], 2))
y_tot_export[:,0] = t_tot
y_tot_export[:,1] = y_tot

np.savetxt('y_raw.csv', y_raw_export, delimiter=',')
np.savetxt('y_tot.csv', y_tot_export, delimiter=',')

# freq derivation by formula = 3.516e-05
# freq derivation by polyfit = 3.513e-05

reference

Jason Sachs. Linear Feedback Shift Registers for the Uninitiated, Part XII: Spread-Spectrum Fundamentals [link]

Jan Meel, Spread Spectrum (SS) — introduction, De Nayer Instituut, Sint-Katelijne-Waver, Belgium, 1999.

Raymond L. Pickholtz, Donald L. Schilling, Laurence B. Milstein, Theory of Spread-Spectrum Communications — A Tutorial, IEEE Transactions on Communications, vol. 30, no. 5, pp. 855-884, May 1982.

Kadeem Samuel. Application Note Clocking for PCIe Applications [https://www.ti.com/lit/an/snaa386/snaa386.pdf?ts=1756864837383]

K. -H. Cheng, C. -L. Hung, C. -H. Chang, Y. -L. Lo, W. -B. Yang and J. -W. Miaw, "A Spread-Spectrum Clock Generator Using Fractional-N PLL Controlled Delta-Sigma Modulator for Serial-ATA III," 2008 11th IEEE Workshop on Design and Diagnostics of Electronic Circuits and Systems, Bratislava, Slovakia, 2008 [https://sci-hub.se/10.1109/DDECS.2008.4538758]

CHUNG-CHUN (CC) CHEN. Why Spread Spectrum Clocking? [https://youtu.be/_PvwTqaWPo8]


Rhee, W. (2020). Phase-locked frequency generation and clocking : architectures and circuits for modern wireless and wireline systems. The Institution of Engineering and Technology

image-20260918203511797

Wang, Zhaowen. Efficient and High-Performance Clocking Circuits for High-Speed Data Links. 2022. Columbia University, PhD dissertation. Academic Commons,[https://academiccommons.columbia.edu/doi/10.7916/g3f1-4e71]

The PI-based architecture decouples the high-frequency clock synthesis and local clock deskew, allowing to optimize the power consumption and circuit area at a system level

Current-Mode Phase Interpolator (CMPI)

A. K. Mishra, Y. Li, P. Agarwal and S. Shekhar, "17.7 A 9b-Linear 14GHz Integrating-Mode Phase Interpolator in 5nm FinFET Process," 2022 IEEE International Solid-State Circuits Conference (ISSCC), San Francisco, CA, USA, 2022, pp. 1-3 [https://sci-hub.jp/10.1109/ISSCC42614.2022.9731703]

—, "Improving Linearity in CMOS Phase Interpolators," in IEEE Journal of Solid-State Circuits, vol. 58, no. 6, pp. 1623-1635, June 2023 [https://sci-net.xyz/10.1109/JSSC.2023.3243305]

D. Xu et al., "8.5 A Scalable Adaptive ADC/DSP-Based 1.25-to-56Gbps/112Gbps High-Speed Transceiver Architecture Using Decision-Directed MMSE CDR in 16nm and 7nm," 2021 IEEE International Solid-State Circuits Conference (ISSCC), San Francisco, CA, USA, 2021, pp. 134-136 [https://sci-hub.ru/10.1109/ISSCC42613.2021.9366063]

Current-mode PIs (CMPIs) can achieve high linearity but at the cost of digital overhead to generate sinusoidal weights

image-20261012235108905

image-20261025195127121

\(\alpha^2 + \beta^2=1\) ensure constant amplitude, however \(\alpha\) still needed be selected with \(\tan\phi = \frac{1-\alpha^2}{\alpha^2}\) for constant phase step

A constant Output amplitude is desired because the swing-dependent delay characteristic of the CML-to-CMOS (C2C) circuit results in AM–PM distortion which eventually manifests as phase nonlinearity

image-20261019235127269

image-20261025180111576

Voltage-Mode Phase Interpolator (VMPI)

A. K. Mishra, Y. Li, P. Agarwal and S. Shekhar, "17.7 A 9b-Linear 14GHz Integrating-Mode Phase Interpolator in 5nm FinFET Process," 2022 IEEE International Solid-State Circuits Conference (ISSCC), San Francisco, CA, USA, 2022, pp. 1-3 [https://sci-hub.jp/10.1109/ISSCC42614.2022.9731703]

image-20261025210925791

image-20261025203937480

image-20261012233712634


B. Razavi, "The Design of a Phase Interpolator [The Analog Mind]," in IEEE Solid-State Circuits Magazine, vol. 15, no. 4, pp. 6-10, Fall 2023 [https://www.seas.ucla.edu/brweb/papers/Journals/BR_SSCM_4_2023.pdf]

Use of resistors in the interpolation network

image-20261019221950869

R High Value Low Value
kink ❎ ✅
nonconformity ✅ ❎

pi_inv_res_iso

With \(\tau_1 = R_1C_1\), \(\tau_2 = R_2C_2\) and \(\tau_{12} = R_1C_2\) \[ \frac{V_o}{V_{in}} = \frac{1}{D(s)}, \quad \frac{V_m}{V_{in}} = \frac{1+s\tau_2}{D(s)} \] where \[ D(s) = 1 + s(\tau_1+\tau_2+\tau_{12}) + s^2\tau_1\tau_2 = (1+s\tau_a)(1+s\tau_b) \]

\[ \tau_{a,b} = \frac{1}{2}\left[(\tau_1+\tau_2+\tau_{12}) \pm \sqrt{(\tau_1+\tau_2+\tau_{12})^2 - 4\tau_1\tau_2}\right], \quad \tau_a > \tau_2 > \tau_b \]

The step responses are \[\begin{align} V_o(t) &= 1 - \frac{\tau_a e^{-t/\tau_a} - \tau_b e^{-t/\tau_b}}{\tau_a - \tau_b} \\ V_m(t) &= 1 - \frac{(\tau_a-\tau_2) e^{-t/\tau_a} + (\tau_2-\tau_b) e^{-t/\tau_b}}{\tau_a - \tau_b} \end{align}\]

  • initial slope: \(V_m'(0^+) = \frac{1}{R_1C_1}\) for any \(R_2\) (\(V_m = V_o = 0\) at \(t=0^+\), no current flows into \(R_2\)), while \(V_o'(0^+) = 0\)
  • Elmore time constant of \(V_m\): \(\int_0^\infty (1-V_m)\,dt = \tau_a + \tau_b - \tau_2 = R_1(C_1+C_2)\), independent of \(R_2\), all of \(C_2\)'s charge still flows through \(R_1\)
  • Elmore time constant of \(V_o\): \(\int_0^\infty (1-V_o)\,dt = \tau_a + \tau_b = R_1C_1 + (R_1+R_2)C_2\), grows with \(R_2\)

For \(R_2 \gg R_1\) (well-separated poles) \[ \boxed{V_m(t) \approx 1 - \frac{R_2}{R_1+R_2}e^{-\frac{t}{(R_1\parallel R_2)C_1}} - \frac{R_1}{R_1+R_2}e^{-\frac{t}{(R_1+R_2)C_2}}} \]

For the internal node (the inverter output), increasing \(R_2\) does not change its Elmore time constant \(R_1(C_1+C_2)\), but it reshapes the response:

  • most of the swing completes with a fast pole \(\approx (R_1\parallel R_2)C_1 \to R_1C_1\), leaving a small tail of amplitude \(\approx \frac{R_1}{R_1+R_2}\) with \(\tau \approx (R_1+R_2)C_2\).

So the inverter edge gets sharper (an earlier mid-rail crossing) even though the output-node time constant \(R_1C_1+(R_1+R_2)C_2\) grows

pi_inv_res_iso_step

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
import numpy as np
import matplotlib.pyplot as plt

R1, C1, C2 = 1, 1, 4 # normalized, t in units of R1*C1
t = np.linspace(0, 25, 25001)

def step(R2):
tau1, tau2, tau12 = R1*C1, R2*C2, R1*C2
b1, b2 = tau1 + tau2 + tau12, tau1*tau2
ta = (b1 + np.sqrt(b1**2 - 4*b2))/2
tb = (b1 - np.sqrt(b1**2 - 4*b2))/2
vo = 1 - (ta*np.exp(-t/ta) - tb*np.exp(-t/tb))/(ta - tb)
vm = 1 - ((ta - tau2)*np.exp(-t/ta) + (tau2 - tb)*np.exp(-t/tb))/(ta - tb)
return vm, vo

plt.figure(figsize=(12, 4.5))
for R2 in [1e-3, 0.3, 1, 3, 10]:
vm, vo = step(R2)
label = r'$R_2 \to 0$' if R2 < 0.01 else rf'$R_2={R2:g}R_1$'
plt.subplot(1, 2, 1); plt.plot(t, vm, linewidth=2, label=label)
plt.subplot(1, 2, 2); plt.plot(t, vo, linewidth=2, label=label)

for i, title in enumerate([r'$V_m$ (inverter output)', r'$V_o$ (shared output)']):
plt.subplot(1, 2, i+1)
plt.axhline(0.5, linestyle='--', color='gray'); plt.grid(); plt.legend(fontsize=10)
plt.title(title, fontsize=14); plt.xlabel(r'$t\,/\,R_1C_1$', fontsize=14)
plt.subplot(1, 2, 1); plt.ylabel('step response', fontsize=14)
plt.suptitle(r'$C_2 = 4C_1$', fontsize=14); plt.tight_layout()

plt.show()

Integrating-Mode Phase Interpolator (IMPI)

A. K. Mishra, Y. Li, P. Agarwal and S. Shekhar, "17.7 A 9b-Linear 14GHz Integrating-Mode Phase Interpolator in 5nm FinFET Process," 2022 IEEE International Solid-State Circuits Conference (ISSCC), San Francisco, CA, USA, 2022, pp. 1-3, doi: 10.1109/ISSCC42614.2022.9731703.

Shaokang ZHAO, 2025, "Multi-Phase Clock Generator for High-Speed Wireline Systems," [paper, slides]

image-20261017142858523

CMPI using LC-resonant load

C. -f. Liang, S. -c. Hwu and S. -i. Liu, "A 10Gbps Burst-Mode CDR Circuit in 0.18μm CMOS," IEEE Custom Integrated Circuits Conference 2006, San Jose, CA, USA, 2006, pp. 599-602 [https://sci-hub.ru/10.1109/CICC.2006.320828]

L. -m. Lee and C. -K. K. Yang, "Phase correction of a resonant clocking system using resonant interpolators," 2008 IEEE Symposium on VLSI Circuits, Honolulu, HI, USA, 2008, pp. 170-171 [https://sci-hub.jp/10.1109/VLSIC.2008.4585994]

M. Erett et al., "A 0.5–16.3 Gbps Multi-Standard Serial Transceiver With 219 mW/Channel in 16-nm FinFET," in IEEE Journal of Solid-State Circuits, vol. 52, no. 7, pp. 1783-1797, July 2017 [https://sci-hub.jp/10.1109/JSSC.2017.2702711]

harmonic filtering for less jitter, better INL & DNL

image-20261025211203236

image-20261025201329931

image-20261025203506679

reference clock edges

B. Razavi, "The Design of a Phase Interpolator [The Analog Mind]," IEEE Solid-State Circuits Magazine, Volume. 15, Issue. 4, pp. 6-10, Fall 2023.(https://www.seas.ucla.edu/brweb/papers/Journals/BR_SSCM_4_2023.pdf)

kink problem

image-20240919223032380

!!! Clock Edges

And for a phase interpolator, you need those reference clocks to be completely the opposite. Ideally they would be triangular shaped

image-20240821203756602

four input clocks given by the cyan, black, magenta, red

Predistortion

Predistortion for sinusoidal

the interpolating inverters near the midscale can be made weaker so as to obtain more uniform phase increments. Alternatively, those at the top and bottom of the array can be made stronger

Single-Quadrant PI

\[ V_o(t) = m \cdot \sin(\omega t + \frac{\pi}{2}) + p\cdot \sin(\omega t) = m\cdot \cos(\omega t) + p\cdot \sin(\omega t) = \sqrt{m^2+p^2} \sin(\omega t + \phi) \]

where \(\tan \phi = \frac{m}{p} = \frac{1-p}{p}\) and \(p = \frac{1}{1+\tan \phi}\)

image-20251016235032393

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
phi = (32:-1:0)./32*pi/2;
p_ideal = 1./(1+tan(phi));
delta_p_ideal = abs(p_ideal(1:end-1) - p_ideal(2:end));
phi_ideal = atan((1-p_ideal)./p_ideal);


p_lin = (0:1:32)/32;
phi_lin = atan((1-p_lin)./p_lin);
delta_p_lin = abs(p_lin(1:end-1) - p_lin(2:end));


delta_plr_predist = ones(1,11)*0.035;
delta_pm_predist = ones(1,10) * (1-2*sum(delta_plr_predist))/10;
delta_p_predist = [delta_plr_predist delta_pm_predist delta_plr_predist];
p_predist = [0 cumsum(delta_p_predist)];
phi_predist = atan((1-p_predist)./p_predist);


subplot(2,2,1)
plot(phi/pi*180, p_ideal, 'ro-', LineWidth=3)
hold on
plot(phi_lin/pi*180, p_lin, 'bs-', LineWidth=3)
grid on; legend('ideal', 'linear', fontsize=12)
xlabel('Phase'); ylabel('p')

subplot(2,2,2)
plot(phi(1:end-1)/pi*180, delta_p_ideal, 'ro-', LineWidth=3)
hold on
plot(phi_lin(1:end-1)/pi*180, delta_p_lin, 'bs-', LineWidth=3)
plot(phi_predist(1:end-1)/pi*180, delta_p_predist, 'gd-', LineWidth=3)
grid on; legend('ideal', 'linear', 'predistortion', fontsize=12)
xlabel('Phase'); ylabel('\Delta p')

subplot(2,2, [3,4])

plot(0:1:32, phi/pi*180, 'ro-', LineWidth=3)
hold on
plot(0:1:32, phi_lin/pi*180, 'bs-', LineWidth=3)
plot(0:1:32, phi_predist/pi*180, 'gd-', LineWidth=3)
grid on; legend('ideal', 'linear', 'predistortion', fontsize=12)
xlabel('code p'); ylabel('Phase')

Eight-Quadrant PI

pi-region.drawio \[\begin{align} V_o(t) &= m \cdot \sin(\omega t + \frac{\pi}{4}) + p\cdot \sin(\omega t) = \frac{\sqrt{2}}{2}m\cdot \cos(\omega t) + \left( \frac{\sqrt{2}}{2}m + p\right)\cdot \sin(\omega t) \\ &= \sqrt{m^2 + p^2 +\sqrt{2}pm}\cdot \sin(\omega t + \phi) \end{align}\]

where \(\tan\phi = \frac{\sqrt{2}m}{\sqrt{2}m+2p} = \frac{\sqrt{2}-\sqrt{2}p}{\sqrt{2}+(2-\sqrt{2})p}\)

image-20251017002836647

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
phi = (16:-1:0)./16*pi/4;
p_ideal = 2^0.5*(1-tan(phi))./(2*tan(phi)+2^0.5*(1-tan(phi)));
delta_p_ideal = abs(p_ideal(1:end-1) - p_ideal(2:end));
phi_ideal = atan((2^0.5 - 2^0.5*p_ideal)./(2^0.5 + (2-2^0.5)*p_ideal));


p_lin = (0:1:16)/16;
phi_lin = atan((2^0.5 - 2^0.5*p_lin)./(2^0.5 + (2-2^0.5)*p_lin));
delta_p_lin = abs(p_lin(1:end-1) - p_lin(2:end));


delta_plr_predist = ones(1,4)*0.066;
delta_pm_predist = ones(1,8) * (1-2*sum(delta_plr_predist))/8;
delta_p_predist = [delta_plr_predist delta_pm_predist delta_plr_predist];
p_predist = [0 cumsum(delta_p_predist)];
phi_predist = atan((2^0.5 - 2^0.5*p_predist)./(2^0.5 + (2-2^0.5)*p_predist));


subplot(2,2,1)
plot(phi/pi*180, p_ideal, 'ro-', LineWidth=3)
hold on
plot(phi_lin/pi*180, p_lin, 'bs-', LineWidth=3)
grid on; legend('ideal', 'linear', fontsize=12)
xlabel('Phase'); ylabel('p')

subplot(2,2,2)
plot(phi(1:end-1)/pi*180, delta_p_ideal, 'ro-', LineWidth=3)
hold on
plot(phi_lin(1:end-1)/pi*180, delta_p_lin, 'bs-', LineWidth=3)
plot(phi_predist(1:end-1)/pi*180, delta_p_predist, 'gd-', LineWidth=3)
grid on; legend('ideal', 'linear', 'predistortion', fontsize=12)
xlabel('Phase'); ylabel('\Delta p')

subplot(2,2, [3,4])

plot(0:1:16, phi/pi*180, 'ro-', LineWidth=3)
hold on
plot(0:1:16, phi_lin/pi*180, 'bs-', LineWidth=3)
plot(0:1:16, phi_predist/pi*180, 'gd-', LineWidth=3)
grid on; legend('ideal', 'linear', 'predistortion', fontsize=12)
xlabel('code p'); ylabel('Phase')

Predistortion for square wave

Weinlader, Daniel, Thomas H. Lee and James A. Gasbarro. "Precision CMOS receivers for VLSI testing applications." (2001). [https://www-vlsi.stanford.edu/people/alum/pdf/0111_Weinlader_Precision_CMOS_Receivers_.pdf]

image-20251017213153657

Normalize Eq. (3.1) with \(\tau = RC\) and threshold at mid-swing, suppose \(V_i(t) = 1- e^{-\frac{t}{\tau}}\) and \(V_q(t) = 1-e^{-\frac{t-\Delta t}{\tau}}\) with \(t\ge \Delta t\) \[ \frac{1}{2} = (1-\alpha)\cdot V_i(t) + \alpha \cdot V_q(t) \] yields the trigger time \[ t = \tau \ln\left[ 1 + \alpha \left(e^{\frac{\Delta t}{\tau}}-1\right)\right] + \tau \ln 2 \] which holds for all \(\alpha\) only if \(\Delta t \le \tau \ln 2\), i.e. the crossing always happens after \(V_q\) starts rising; otherwise, at small \(\alpha\) the output crosses before \(V_q\) arrives → kink problem

Then \[\begin{align} \frac{\partial t}{\partial \alpha} &= \tau \frac{e^{\frac{\Delta t}{\tau }}-1}{1+\alpha(e^{\frac{\Delta t}{\tau}}-1)} \gt 0 \\ \frac{\partial^2 t}{\partial \alpha^2} &= -\tau \frac{\left(e^{\frac{\Delta t}{\tau }}-1\right)^2}{\left(1+\alpha(e^{\frac{\Delta t}{\tau}}-1)\right)^2} \lt 0 \end{align}\]

  • full span: \(t(1) - t(0) = \Delta t\)
  • step ratio: \(\left.\frac{\partial t}{\partial \alpha}\right|_{\alpha=0} \Big/ \left.\frac{\partial t}{\partial \alpha}\right|_{\alpha=1} = e^{\frac{\Delta t}{\tau}}\)
  • peak INL \(\approx \frac{\Delta t^2}{8\tau}\) at \(\alpha \approx \frac{1}{2}\)

Hence, larger weight step \(\Delta \alpha\) as \(\alpha\) approaches 1 in order to improve linearity. Unlike the sinusoidal case (symmetric, S-shaped), the correction is one-sided

For \(N\) uniform delay steps \(t_k - t_0 = \frac{k}{N}\Delta t\), the predistorted weight is \[ \alpha_k = \frac{e^{\frac{k}{N}\frac{\Delta t}{\tau}}-1}{e^{\frac{\Delta t}{\tau}}-1}, \quad k = 0, 1, \dots, N \]

image-20261013001122226

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
import numpy as np
import matplotlib.pyplot as plt

tau = 15 # ps
dt = 6 # ps, \Delta t
N = 100 # number of codes
t = np.linspace(dt, 56, 500001)

vi = 1 - np.exp(-t/tau)
vq = 1 - np.exp(-(t-dt)/tau)

def t_cross(alpha):
viq = (1-alpha)*vi + alpha*vq
return t[np.argmin(np.abs(viq - 0.5))]

k = np.arange(N+1)
alpha_lin = k/N
alpha_predist = (np.exp(k/N*dt/tau) - 1)/(np.exp(dt/tau) - 1)

td_lin = np.array([t_cross(a) for a in alpha_lin])
td_lin = td_lin - td_lin[0]
td_predist = np.array([t_cross(a) for a in alpha_predist])
td_predist = td_predist - td_predist[0]
td_closed = tau*np.log(1 + alpha_lin*(np.exp(dt/tau) - 1))

plt.figure(figsize=(16,10))
plt.subplot(2, 2, 1)
for alpha in alpha_lin:
plt.plot(t[::1000], ((1-alpha)*vi + alpha*vq)[::1000])
plt.plot([0, 60], [0.5,0.5], '--c', linewidth=3); plt.grid()
plt.xlabel('t', fontsize=14); plt.ylabel('Voltage', fontsize=14)

plt.subplot(2, 2, 2)
plt.plot(k, alpha_lin, 'bs-', label='linear')
plt.plot(k, alpha_predist, 'gd-', label='predistortion')
plt.grid(); plt.legend(fontsize=12)
plt.xlabel('code k', fontsize=14); plt.ylabel(r'$\alpha$', fontsize=14)

plt.subplot(2, 2, 3)
plt.plot(k, td_lin, 'bs-', label='linear')
plt.plot(k, td_closed, 'r--', linewidth=3, label='closed form', zorder=3)
plt.plot(k, td_predist, 'gd-', label='predistortion')
plt.grid(); plt.legend(fontsize=12)
plt.xlabel('code k', fontsize=14); plt.ylabel(r'$t_d$', fontsize=14)

plt.subplot(2, 2, 4)
plt.plot(k[:-1], np.diff(td_lin), 'bs-', label='linear')
plt.plot(k[:-1], np.diff(td_predist), 'gd-', label='predistortion')
plt.grid(); plt.legend(fontsize=12)
plt.xlabel('code k', fontsize=14); plt.ylabel(r'$\Delta t_d$', fontsize=14)

plt.show()

PI static & dynamic linearity

A. K. Mishra, Y. Li, P. Agarwal and S. Shekhar, "17.7 A 9b-Linear 14GHz Integrating-Mode Phase Interpolator in 5nm FinFET Process," 2022 IEEE International Solid-State Circuits Conference (ISSCC), San Francisco, CA, USA, 2022, pp. 1-3, [https://sci-hub.jp/10.1109/ISSCC42614.2022.9731703]

—, "Improving Linearity in CMOS Phase Interpolators," in IEEE Journal of Solid-State Circuits, vol. 58, no. 6, pp. 1623-1635, June 2023, [https://sci-net.xyz/10.1109/JSSC.2023.3243305]

Wang, Zhaowen. Efficient and High-Performance Clocking Circuits for High-Speed Data Links. 2022. Columbia University, PhD dissertation. Academic Commons,[https://academiccommons.columbia.edu/doi/10.7916/g3f1-4e71]

PI rotation (dynamic operation)

image-20260919083229868

\[ K_{f,PI} = \frac{1/2^N}{T_m/T_o}\cdot \frac{1}{T_o} = \frac{f_m}{2^N} \] pi-code.drawio


For DCO \[ K_{f,DCO} = \frac{K_T}{T_o}\cdot \frac{1}{T_o} = \frac{K_T}{T_o^2} \]


image-2026-10-13_12-46

Deterministic Jitter & rotation spur

E. Monaco, G. Anzalone, G. Albasini, S. Erba, M. Bassi and A. Mazzanti, "A 2–11 GHz 7-Bit High-Linearity Phase Rotator Based on Wideband Injection-Locking Multi-Phase Generation for High-Speed Serial Links in 28-nm CMOS FDSOI," in IEEE Journal of Solid-State Circuits, vol. 52, no. 7, pp. 1739-1752, July 2017 [https://sci-hub.jp/10.1109/JSSC.2017.2702742]

image-20261025184806969 \[ \boxed{\color{red}|\text{DJ}| \leq T_{LSB} (1+\max(\text{DNL}) + \text{INL}_\text{pk-pk})} \]


image-20260919082930453

Let \(T=T_{\mathrm{LSB}}\), and write a PI’s output time as

\[ t(k)=t_{\mathrm{ref}}+[k+I(k)]T \]

where \(I(k)=\mathrm{INL}(k)\), expressed in LSBs. Then

\[ \mathrm{DNL}(k)=\frac{t(k+1)-t(k)}{T}-1 =I(k+1)-I(k) \]

Thus INL describes the error at a code, while DNL describes the error in one code-to-code step.

1. Why \((0.5+|\mathrm{DNL}_p|)T\)?

For an ideal PI, available phases are spaced by \(T\). If the desired transition lies halfway between two available phases, selecting the nearest phase leaves an error of

\[ |E_e|\le \frac{T}{2} \]

That explains the \(0.5\)

With DNL, a particular step has width

\[ t(k+1)-t(k)=[1+\mathrm{DNL}(k)]T \]

A larger step means a larger gap in which the desired phase might lie.

The excerpt can be read as budgeting

\[ \underbrace{0.5T}_{\text{nominal quantization}} + \underbrace{|\mathrm{DNL}_p|T}_{\text{nonlinearity allowance}} \]

But for a monotonic PI that actually selects the nearest available phase, the tighter bound is half the largest actual step:

\[ \boxed{|E_e| \le \frac{1+\max_k\mathrm{DNL}(k)}{2}T \le \left(0.5+\frac{|\mathrm{DNL}_p|}{2}\right)T} \]

So the paper's \(0.5+|\mathrm{DNL}_p|\) is a looser bound under this model

2. Why does the data-clock bound become \((1+|\mathrm{DNL}_p|+|\mathrm{INL}_{pp}|)T\)?

\(E_e\) and \(E_d\) are timing errors, measured in seconds—not the clock times themselves

  • \(E_e\): edge-sampling clock error relative to the ideal data-transition time: \[ E_e=t_e-t_{\text{transition}}. \]

  • \(E_d\): data-sampling clock error relative to the ideal data-sampling time, assumed here to be half a UI after that transition: \[ E_d=t_d-\left(t_{\text{transition}}+\frac{\mathrm{UI}}{2}\right). \]

Here \(t_e\) and \(t_d\) are the actual sampling times. A positive error means the clock samples late; a negative error means it samples early.

Writing \(H=\mathrm{UI}/2\), these definitions give

\[ \boxed{E_d=E_e+\underbrace{(t_d-t_e-H)}_{\text{error in edge-to-data spacing}}.} \]

Let the desired edge-to-data spacing be

\[ H=\frac{\mathrm{UI}}{2} \]

and let the data-clock code be \(k+m\) when the edge-clock code is \(k\). Then

\[ t_d-t_e=mT+[I(k+m)-I(k)]T \]

Consequently, the data-clock error relative to its ideal sampling position is

\[ \boxed{ E_d = E_e +\underbrace{(mT-H)}_{\text{spacing quantization}} +\underbrace{[I(k+m)-I(k)]T}_{\text{relative INL error}} } \]

If \(m\) is chosen by rounding \(H/T\), then

\[ |mT-H|\le 0.5T. \]

Combining this with the paper’s edge-clock allowance gives

\[ |E_d| \le \underbrace{(0.5+|\mathrm{DNL}_p|)T}_{\text{edge-clock error}} +\underbrace{0.5T}_{\text{spacing quantization}} +\underbrace{\mathrm{INL}_{pp}T}_{\text{relative INL error}} \]

which produces the quoted expression.

Using tighter edge-clock bound

\[ |E_d| \le \underbrace{\left(0.5+\frac{|\mathrm{DNL}_p|}{2}\right)T_{\mathrm{LSB}}}_{\text{edge-clock error}} +\underbrace{0.5T_{\mathrm{LSB}}}_{\text{spacing quantization}} +\underbrace{\mathrm{INL}_{pp}T_{\mathrm{LSB}}}_{\text{relative INL error}} \]

so

\[ \boxed{|E_d|\le \left(1+\frac{|\mathrm{DNL}_p|}{2}+\mathrm{INL}_{pp}\right)T_{\mathrm{LSB}}} \]

If \(H/T\) is an integer—for example, a full-period PI with a number of steps divisible by eight can represent \(45^\circ\) exactly—then \(mT-H=0\). That extra \(0.5T\) is unnecessary. (If the desired edge-to-data spacing is exactly representable by an integer number of PI steps, the spacing-quantization term vanishes, and the constant \(1\) becomes \(0.5\))


The CDR finds an edge-clock code, and the data-clock code is obtained by adding a fixed code offset

With \(T=T_{\mathrm{LSB}}\) and \(H=\mathrm{UI}/2\):

\[ k_e=k,\qquad m=\operatorname{round}\left(\frac{H}{T}\right),\qquad k_d=k_e+m \]

Here, \(m\) is a number of PI steps, not a time. Thus, for an ideal PI, the timing relationship is

\[ \boxed{ t_d=t_e+\operatorname{round}\left(\frac{\mathrm{UI}}{2T}\right)T} \]

The circuit operates continuously: the CDR adjusts \(k_e\), and the data-clock code follows as \(k_d=k_e+m\). It does not need to measure a numerical value of \(t_e\) before generating the data clock.

For nonlinear PIs, adding \(m\) codes does not necessarily add exactly \(mT\) in time. Assuming a common timing reference,

\[ \boxed{ t_d=t_e+mT+ \left[I_d(k_e+m)-I_e(k_e)\right]T} \]

That last term is precisely the relative INL error we discussed. The subscripts allow the edge and data clocks to come from different PIs.

PI vs. PLL based CDR

PCI Express Jitter Modeling Revision 1.0RD July 14, 2004

image-20250816121744921

image-20260602202522018 \[ H_1 - \left[H_1(1-H_3) + H_2H_3\right] = (H_1-H_2)H_3 \]

reference

Cortiula A, Menin D, Bandiziol A, Driussi F, Palestri P. Modeling of Phase-Interpolator-Based Clock and Data Recovery for High-Speed PAM-4 Serial Interfaces. Electronics. 2025; [https://www.mdpi.com/2079-9292/14/10/1979]

G. Souliotis, A. Tsimpos and S. Vlassis, "Phase Interpolator-Based Clock and Data Recovery With Jitter Optimization," in IEEE Open Journal of Circuits and Systems, vol. 4, pp. 203-217, 2023 [https://ieeexplore.ieee.org/document/10184121]

T. Chan Carusone, T. O. Dickson, S. Palermo, S. Shekhar and M. Mansuri, "Modern Wireline Transceivers," in IEEE Journal of Solid-State Circuits, vol. 61, no. 2, pp. 395-422, Feb. 2026 [https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=11311714]

CHUNG-CHUN (CC) CHEN, Why Phase Interpolator Based CDR? [https://youtu.be/FkoJwnAYbMs]

—, Why High Linearity Phase Interpolator with Low Power? [https://youtu.be/II9vQPH8MNM]

—, Why Not A Voltage Mode or A Current Mode Phase Interpolator? [https://youtu.be/1vrBgDNTOMA]

—, Why An Integrating Mode Phase Interpolator? [https://youtu.be/VEkM8ZEcYnc]

—, Why High-Speed Integrating Mode Phase Interpolator, IMPI? [https://youtu.be/6197voGpOek]

rfinsights, phase rotator in serdes - every major topology, explained [https://www.rfinsights.com/synthesizer/phase-rotator/]

image-20250814200158666

image-20250814185129279

image-20251211223949724

image-20251211224346954

MTBF Calculation

image-20251210215011221

cassync_tr.drawio

image-20251210213433116


Beer, Salomon & Priel, Michael & Dobkin, Rostislav & Kolodny, Avinoam. (2010). The Devolution of Synchronizers. Proceedings - International Symposium on Asynchronous Circuits and Systems. [pdf]

\(T_W\), metastability window is defined differently among published paper

image-20251209232205525

Some PDK provide \(T_0\) and \(\tau\) for the corresponding cascaded flip flop synchronizer stdcells \[ \text{MTBF} = \frac{1}{T_W\times f_C\times f_D}\space \space \text{,where}\space\space T_W = T_0 e^{-T_r/\tau} \] and \(T_r = T_C - T_{DLY} - T_{SU}\)

Ran Ginosar

ISCAS 2008 tutorial: Synchronization Circuits for Multiple Clock Domains SoCs

Enter metastabilty

image-20260402210557862



Exit metastabilty

image-20260402211731588



MTBF (Mean Time Between Failures)

image-20260402211924586

image-20260402212255182

image-20260402212425616

image-20260402212648758

image-20260402211215546



MTBF of Many Synchronizers

image-20260402211014731

Synchronizer Characterization

I. W. Jones, S. Yang and M. Greenstreet, "Synchronizer Behavior and Analysis," 2009 15th IEEE Symposium on Asynchronous Circuits and Systems, Chapel Hill, NC, USA, 2009 [https://sci-hub.ru/10.1109/ASYNC.2009.8]

Xprova. bisect-tau - EDA tool for characterizing the metastability resolution time constant (Tau) of bistable circuits [https://github.com/xprova/bisect-tau]

For GNU Octave, version 8.4.0, ngspice-42 : Circuit level simulation program @Ubuntu 24.04.3 LTS x86_64

[https://github.com/raytroop/bisect-tau]

TODO 📅

image-20251211232230119

DFF Simulation

image-20250815012602436

The typical flip-flops comprise master and slave latches and decoupling inverters.

In metastability, the voltage levels of nodes A and B of the master latch are roughly midway between logic 1 (VDD) and 0 (GND)

master latch enter metastability

In fact, one popular definition says that if the output of a flip-flop changes later than the nominal clock-to-Q propagation delay, then the flip-flop must have been metastable


sweep \(\Delta t_{D \to \space \text{CK}}\)

image-20250815181257083


transient noise analysis @ \(\Delta t_{D \to \space \text{CK}} = -3.444525p\)

image-20250815190341687

zoom out

image-20250815190431960


image-20250815011210280

Noise Seed—Seed for the random number generator (used by the simulator to vary the noise sources internally). Specifying the same seed allows you to reproduce a previous experiment. The default value is 1.

Synchronizer effect – latency uncertainty

image-20250814202542548

Reset Synchronizer

[https://vlsiuniverse.blogspot.com/2016/09/reset-synchronizer.html]

Clifford E. Cummings, Asynchronous & Synchronous Reset Design Techniques - Part Deux [http://staff.ustc.edu.cn/~wyu0725/FPGA/snug_collection/2SNUG%20Design/Asynchronous%20and%20Synchronous%20Reset%20Design%20Techniques_ppt.pdf]

The way most of the designs have been modelled needs asynchronous reset assertion and synchronous de-assertion

img

image-20261113111551243

Data Synchronizer

[https://vlsiuniverse.blogspot.com/2013/09/synchronization-schemes.html]

reset_data_synchronizer.drawio

using rst_sync_n to reset all three data synchronizer stages is a standard, reasonable arrangement, provided both synchronizers use the same local_clk. It implements asynchronous reset assertion and synchronous reset deassertion

reference

Yvain Thonnart, CEA-LIST. ISSCC2021 T8: On-Chip Interconnects: Basic Concepts, Designs and Future Opportunities

R. Ginosar, "Metastability and Synchronizers: A Tutorial," in IEEE Design & Test of Computers, vol. 28, no. 5, pp. 23-35, Sept.-Oct. 2011 [https://webee.technion.ac.il/~ran/papers/Metastability-and-Synchronizers.IEEEDToct2011.pdf] [color]

Steve Golson. Synchronization and Metastability [https://trilobyte.com/pdf/golson_snug14.pdf]

Kinniment, D. J. Synchronization and arbitration in digital systems. John Wiley & Sons Ltd (2007).

Synchronizers And Data FlipFlops are Different [https://ee.usc.edu/async2015/web/wp-content/uploads/2015/03/S1_P4_ASYNC2015IndustrialPaperDFF.pdf]

S. Beer, R. Ginosar, M. Priel, R. Dobkin and A. Kolodny, "The Devolution of Synchronizers," 2010 IEEE Symposium on Asynchronous Circuits and Systems, Grenoble, France, 2010 [pdf]

赵启林 klin, Metastability

Asad Abidi. ISSCC 2023: Circuit Insights "The CMOS Latch" [https://youtu.be/sVe3VUTNb4Q&t=681]

Matt Venn. Inside a flip-flop: exploring metastability [https://zerotoasiccourse.com/metastability/] [https://github.com/mattvenn/flipflop_demo]

Amr Adel Mohammady. Clock Domain Crossing [linkedin]

A dynamical system can be linear or nonlinear. Independently, it can be deterministic or stochastic. Continuous-time deterministic systems are commonly modeled by ODEs, while continuous-time stochastic systems are commonly modeled by SDEs

Deterministic Stochastic
Linear Linear ODE Linear SDE
Nonlinear Nonlinear ODE Nonlinear SDE

The two classifications answer different questions:

  • Linear/nonlinear: How does the state enter the evolution equation?
  • Deterministic/stochastic: Does the evolution include randomness?

For Demir’s oscillator theory, however, the main path is \[ \boxed{ \text{nonlinear deterministic ODE} \rightarrow \text{add device noise} \rightarrow \text{nonlinear SDE} } \]

instantaneous & average PSD

Darabi H. Radio Frequency Integrated Circuits and Systems. 2nd ed. Cambridge University Press; 2020.

For white noise \(n(t)\)

image-20260725005940162

flicker noise Modulation

flicker noise spectrum

Darabi H. Radio Frequency Integrated Circuits and Systems. 2nd ed. Cambridge University Press; 2020.

P. Yelleswarapu, A. Jha, R. Willis, Y. Makris and K. K. O, "Phase Noise Reduction in LC VCO’s Using an Array of Cross-Coupled Nanoscale MOSFETs and Intelligent Post-Fabrication Selection," in IEEE Transactions on Microwave Theory and Techniques, vol. 70, no. 6, pp. 3244-3256, June 2022 [https://sci-hub.ru/10.1109/TMTT.2022.3164949]

image-20261017092235495

image-20260724230315724

image-20260724230516271

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
f       = logspace(0, 10, 4000);      % 1 Hz ... 10 GHz
tau_min = 1e-9; % fastest trap (corner ~160 MHz)
tau_max = 1e-2; % slowest trap (corner ~16 Hz)
r = tau_max/tau_min; % 7 decades of time constants
Nlist = [1 3 300]; % number of superposed traps

figure('Color','w','Position',[100 60 720 800]);

% ------------------------------ spectra ------------------------------
ax1 = subplot(2,1,1); hold(ax1,'on');
for m = 1:numel(Nlist)
N = Nlist(m);
if N == 1
tau = sqrt(tau_min*tau_max); % mid-band trap
else
tau = logspace(log10(tau_min), log10(tau_max), N); % log-spaced
end
S = zeros(size(f));
for k = 1:numel(tau)
S = S + tau(k) ./ (1 + (2*pi*f*tau(k)).^2); % c_t = 1
end
% Every Lorentzian carries the same total power (integral over f = 1/4
% regardless of tau), so dividing by N keeps the total variance fixed:
S = S/N;
plot(ax1, f, S, 'LineWidth', 2, 'DisplayName', sprintf('N = %d', N));
end

xline(ax1, 1/(2*pi*tau_max), 'Color', [.85 .85 .85], 'LineWidth',2, 'HandleVisibility', 'off');
xline(ax1, 1/(2*pi*tau_min), 'Color', [.85 .85 .85], 'LineWidth',2, 'HandleVisibility', 'off');
set(ax1, 'XScale','log', 'YScale','log'); grid(ax1,'on');
xlim(ax1, [f(1) f(end)]);
xlabel(ax1, 'f (Hz)'); ylabel(ax1, 'S(f) / (c_t N) (a.u.)');
title(ax1, 'Superposition of trap Lorentzians \rightarrow 1/f');
legend(ax1, 'Location', 'southwest', 'FontSize', 8);

% --------------------------- local slope -----------------------------
ax2 = subplot(2,1,2); hold(ax2,'on');
for m = 1:numel(Nlist)
N = Nlist(m);
if N == 1
tau = sqrt(tau_min*tau_max);
else
tau = logspace(log10(tau_min), log10(tau_max), N);
end
S = zeros(size(f));
for k = 1:numel(tau)
S = S + tau(k) ./ (1 + (2*pi*f*tau(k)).^2);
end
plot(ax2, f, gradient(log(S))./gradient(log(f)), 'LineWidth', 2);
end
% plot(ax2, f, gradient(log(SL))./gradient(log(f)), '-.', 'Color', [.55 .55 .55]);
yline(ax2, -1, ':', '1/f', 'Color', [.85 .1 .2], 'LineWidth', 1.2);
yline(ax2, -2, ':', '1/f^2', 'Color', 'k');
xline(ax2, 1/(2*pi*tau_max), 'Color', [.85 .85 .85], 'LineWidth', 2);
xline(ax2, 1/(2*pi*tau_min), 'Color', [.85 .85 .85], 'LineWidth', 2);
set(ax2, 'XScale', 'log'); grid(ax2,'on');
xlim(ax2, [f(1) f(end)]); ylim(ax2, [-2.4 0.25]);
xlabel(ax2, 'f (Hz)'); ylabel(ax2, 'd logS / d logf');
title(ax2, 'Local log-log slope: one trap \rightarrow -2, many traps \rightarrow -1');

numerical generation of flicker noise

Bibbona, Enrico, Gianna Panfilo and Patrizia Tavella. "The Ornstein–Uhlenbeck process as a model of a low pass filtered white noise." Metrologia 45 (2008): S117 - S126. [https://iris.polito.it/retrieve/e384c42f-3847-d4b2-e053-9f05fe0a1d67/OUasFWN_finale.pdf]

Ornstein–Uhlenbeck process, equivalently white noise passed through a first-order low-pass filter

image-20260801203350117

image-20260801203516575

Flicker Noise Formulations in Verilog-A

G. J. Coram, C. C. McAndrew, K. K. Gullapalli and K. S. Kundert, "Flicker Noise Formulations in Compact Models," in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 39, no. 10, pp. 2812-2821, Oct. 2020 [https://kenkundert.com/docs/tcad20-flicker-noise.pdf],[https://github.com/KenKundert/flicker-noise]

BSIM4v4.7 MOSFET Model -User's Manual [https://class.ece.iastate.edu/djchen/ee501/BSIM470_Manual.pdf]

C. C. McAndrew et al., "Best Practices for Compact Modeling in Verilog-A," in IEEE Journal of the Electron Devices Society, vol. 3, no. 5, pp. 383-396, Sept. 2015 [https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=7154394]

image-20260801190953551

When sign(Ir) = -1, the argument becomes \[ q(t)=\operatorname{sign}(I_r)P_n=-P_n. \] A simulator that correctly supports Kundert’s formulation does not interpret this as a physically negative PSD, nor does it calculate the ordinary complex square root \(\sqrt{-P_n}\). Instead, the sign selects the sign of the deterministic noise-modulation amplitude: \[ \boxed{ m(t)=\operatorname{sign}\!\big(q(t)\big)\sqrt{|q(t)|} } \] Therefore, when \(q=-P_n\), \[ m(t)=-\sqrt{P_n}. \] This is equivalent to

1
I(a,b) <+ sign(Ir)*flicker_noise(Pn, EF, "flicker");

provided the simulator supports a noise function inside an expression

image-20260802114438747

flicker_noise_commutation_vs_abs_static

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
// BSIM flicker noise simulations

simulator lang=spectre

model nchbsim4_f0 bsim4 fnoimod=0 kf=1e-23 af=2
model nchbsim4_f1 bsim4 fnoimod=1

Vmod (mod 0) vsource type=sine dc=1.0 sinedc=0.0 ampl=100mV freq=131.072kHz
ED (d 0 mod 0) vcvs gain=1
ES (s 0 mod 0) vcvs gain=-1
VG (g 0) vsource dc=3
VB (b 0) vsource dc=-0.2

MBSIM4f0 (d_f0 g s b) nchbsim4_f0 l=1um w=10um
MBSIM4f1 (d_f1 g s b) nchbsim4_f1 l=1um w=10um

iRESf0 (d d_f0) vsource dc=0.0
iRESf1 (d d_f1) vsource dc=0.0
Rout (noise 0) resistor isnoisy=no r=100kOhm
Hnoise (noise 0) pccvs coeffs=[0 1 1] probes=[iRESf0 iRESf1]

noise (noise 0) noise start=4_Hz stop=4.194304MHz dec=2k
pop pss fund=131.072kHz
pnoise (noise 0) pnoise start=4_Hz stop=4.194304MHz dec=2k maxsideband=10
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
// Resistor flicker noise simulations

simulator lang=spectre

ahdl_include "resistor.va"

model rref resistor kf=1.0e-6 af=2 // to match res_va

Vmod (n 0) vsource type=sine dc=1.0 sinedc=0.0 ampl=100mV freq=131.072kHz
Rva (n 0) res_va
Rref (n 0) rref r=100.0

noise noise start=4_Hz stop=4.194304MHz dec=2k oprobe=Vmod
pss pss fund=131.072kHz maxacfreq=4.194304MHz
pnoise pnoise start=4_Hz stop=4.194304MHz dec=2k maxsideband=10 oprobe=Vmod

Marek Mierzwinski, Verilog-A Standardization for Compact Modeling [https://www.mos-ak.org/washington_dc/papers/Mierzwinski_MOS-AK_2011.pdf]

Current BSIM models use compact-model equations standardized through reference Verilog-A code, but commercial simulators often execute an optimized built-in implementation rather than the Verilog-A source directly

image-20260729232946151

image-20260729233151017

image-20260729233304135

flicker noise in circuit-noise analysis

its power spectral density is approximately \[ S_{i,1/f}(f)=\frac{K}{|f|}. \] A large amount of its power lies at low frequencies. Therefore, compared with a GHz oscillation period \(T_0\), the flicker-noise value changes very little during one cycle.

For a flicker-noise component at frequency \(f_m\), \[ f_m T_0\ll 1 \] implies \[ i_{1/f}(t+T_0)\approx i_{1/f}(t). \] Thus, if the noise current is positive at \(t_0\), it will probably remain positive throughout the following oscillator cycle: \[ i_{1/f}(t_0+\tau)\approx i_{1/f}(t_0), \qquad 0\leq \tau<T_0. \] In circuit-noise analysis, the underlying flicker-noise source is commonly treated as approximately wide-sense stationary: \[ R_x(t_1,t_2)=R_x(t_1-t_2). \] This is reasonable when the device bias is constant and the measurement interval is finite.

The phase perturbations may cancel or leave a nonzero residual: \[ \Delta\phi_{\text{cycle}} \propto \int_{0}^{T_0} \Gamma(\omega_0 t)\, i_{1/f,\mathrm{cyclo}}(t)\,dt. \] Since the low-frequency noise is almost constant over \(T_0\), \[ \Delta\phi_{\text{cycle}} \approx x_{1/f}(t_0) \int_{0}^{T_0} \Gamma(\omega_0 t)a(t)\,dt \] Therefore, flicker-noise upconversion depends on whether the phase-delay and phase-advance contributions cancel over one period. A nonzero weighted average produces low-frequency fluctuations in oscillator frequency, which commonly appear as the \(1/f^3\) phase-noise region.

Define

\[ \Gamma_{\mathrm{eff,DC}}\equiv \frac{1}{T_0}\int_0^{T_0}\Gamma(\omega_0t)a(t)\,dt \]

Then \[ \Delta\phi_{\text{cycle}} \approx \frac{x_{1/f}(t_0)}{q_{\max}} \Gamma_{\mathrm{eff,DC}}T_0. \] If \(x_{1/f}\) is already normalized by \(q_{\max}\), the \(1/q_{\max}\) factor can be omitted.

Therefore, \[ \boxed{\Gamma_{\mathrm{eff,DC}}=0 \quad\Longrightarrow\quad \Delta\phi_{\text{cycle}}\approx 0} \] for quasistatic flicker noise. Physically, the phase-delay contribution on one edge exactly cancels the phase-advance contribution on the other edge.nce, \[ \boxed{ \Gamma_{\mathrm{eff,DC}}=0 \Rightarrow \text{no first-order direct }1/f\text{-to-}1/f^3 \text{ phase-noise upconversion from that source.} } \]

Ordinary Differential Equations (ODEs)

Steve Brunton, ME 564 - Mechanical Engineering Analysis [http://faculty.washington.edu/sbrunton/me564/] [videos]

Dirac delta function in ODEs

Integrate across the impulse to find the jump \[ \underbrace{\text{zero state} + \delta(t)\text{ input}}_{t=0^-} \quad\Longrightarrow\quad \underbrace{\text{zero input} + \text{new ICs at }t=0^+}_{t>0} \]

image-20260710230815001

image-20260711003430785

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
import numpy as np
import scipy.integrate as spi
import matplotlib.pyplot as plt


L, C, R = 2.533e-9, 10e-12, 100.0

def rhs(t, y):
il , dil = y
dildt = dil
ddildt = -(1/(R*C))*dil - (1/(L*C))*il
return [dildt, ddildt]

sol = spi.solve_ivp(rhs, (0, 10e-9), [0, 1/(L*C)],
t_eval=np.linspace(0, 10e-9, 2001),
rtol=1e-10, atol=1e-4) # error drops to ~1e-10

plt.plot(sol.t, sol.y[0])
plt.title('RLC Circuit Response')
plt.xlabel('Time (s)')
plt.ylabel('Current (A)')
plt.grid()
plt.show()

image-20260710235118022

Doublet function in ODEs

image-20260711004412242

image-20260711004554172

image-20260711010924471

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
import numpy as np
import scipy.integrate as spi
import matplotlib.pyplot as plt


RC = 1.0
T_START = 0.0
T_STOP = 20.0
NUM_SAMPLES = 2001


def wien_bridge_rhs(t, state):
"""Return the state derivative for the normalized Wien bridge response."""
#del t
vo, dvo = state
ddvo = -(3.0 / RC) * dvo - vo / RC**2
return [dvo, ddvo]


def analytic_response(t):
"""Closed-form voltage response for the same initial conditions."""
sqrt_5 = np.sqrt(5.0)
s1 = (-3.0 + sqrt_5) / (2.0 * RC)
s2 = (-3.0 - sqrt_5) / (2.0 * RC)
return (s1 * np.exp(s1 * t) - s2 * np.exp(s2 * t)) / sqrt_5


def main():
t_eval = np.linspace(T_START, T_STOP, NUM_SAMPLES)
initial_state = [1.0 / RC, -3.0 / RC**2]

sol = spi.solve_ivp(
wien_bridge_rhs,
(T_START, T_STOP),
initial_state,
t_eval=t_eval,
rtol=1e-10,
atol=1e-8,
)

fig, ax = plt.subplots(figsize=(8, 4.8), constrained_layout=True)
ax.plot(sol.t, sol.y[0], linewidth=3.0, label="Numerical solution")
ax.plot(t_eval, analytic_response(t_eval), "--", linewidth=3.0, label="Analytic response")

ax.set_title("Wien Bridge Natural Response")
ax.set_xlabel("Time (s)")
ax.set_ylabel("Output voltage, $v_o$ (V)")
ax.grid(True, which="both", linestyle=":", linewidth=0.8, alpha=0.8)
ax.legend(frameon=False)

plt.show()


if __name__ == "__main__":
main()

Stochastic Differential Equations (SDE)

TODO 📅

Fourier Analysis & Partial Differential Equations (PDEs)

\[ \text{Fourier analysis} \longrightarrow \text{method for solving PDEs}, \]

TODO 📅

Differential Equations in Matlab & Python

scipy.integrate.solve_ivp

Solve an initial value problem for a system of ODEs

rtol and atol are the error tolerances for scipy.integrate.solve_ivp.

rtol is relative tolerance: allowed error scales with the size of the solution.

atol is absolute tolerance: allowed error floor when the solution is near zero.

SciPy roughly controls local error using:

1
error < atol + rtol * abs(y)

DifferentialEquations.jl

ODE forms

ODE is usually defined in one of two forms: out-of-place or in-place

1
2
3
4
5
6
7
8
9
10
11
# out-of-place

function f(x, p, t)
return 2x
end

x0 = 1.0
tspan = (0.0, 5.0)

prob = ODEProblem(f, x0, tspan)
sol = solve(prob)

x = current state

p = parameters

t = current time

returned value = dx/dt

1
2
3
4
5
6
7
8
9
10
11
# in-place

function f!(dx, x, p, t)
dx[1] = 2*x[1] # calars cannot be mutated
end

x0 = [1.0]
tspan = (0.0, 5.0)

prob = ODEProblem(f!, x0, tspan)
sol = solve(prob)

Scalar ODE \[ \frac{dx}{dt} = -2x \]

1
2
3
4
5
6
7
8
9
function f(x, p, t)
return 2x
end

x0 = 1.0
tspan = (0.0, 5.0)

prob = ODEProblem(f, x0, tspan)
sol = solve(prob)

System of ODEs

\[\begin{align} \dot{x} &= y, \\ \dot{y} &= -x - 0.2y. \end{align}\]

1
2
3
4
5
6
7
8
9
10
11
12
13
function oscillator!(du, u, p, t)
x = u[1]
y = u[2]

du[1] = y
du[2] = -x - 0.2y
end

u0 = [1.0, 0.0]
tspan = (0.0, 20.0)

prob = ODEProblem(oscillator!, u0, tspan)
sol = solve(prob)

\[ u = \begin{bmatrix} x \\ y \end{bmatrix}, \quad du = \begin{bmatrix} \dot{x} \\ \dot{y} \end{bmatrix}. \]

ODE with parameters

1
2
3
4
5
6
7
8
9
10
11
12
13
function f!(du, u, p, t)
a, b = p
x = u[1]

du[1] = a*x - b*x^3
end

u0 = [0.1]
p = (2.0, 1.0)
tspan = (0.0, 10.0)

prob = ODEProblem(f!, u0, tspan, p)
sol = solve(prob)

The Lorenz Equation — employ above features

\[\begin{align} \frac{dx}{dt} &= \sigma (y - x) \\ \frac{dy}{dt} &= x (\rho - z) -y \\ \frac{dz}{dt} &= xy - \beta z \end{align}\]

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
function lorenz!(du,u,p,t)
σ,ρ,β = p
du[1] = σ*(u[2]-u[1])
du[2] = u[1]*(ρ-u[3]) - u[2]
du[3] = u[1]*u[2] - β*u[3]
end

u0 = [1.0,0.0,0.0]
p = (10,28,8/3) # we could also make this an array, or any other type!
tspan = (0.0,100.0)

prob = ODEProblem(lorenz!,u0,tspan,p)
sol = solve(prob)

Plots.plot(sol, vars=(1,2,3), size=(1400, 700))

image-20260815142842589

Event Handling & Callback Functions

In DifferentialEquations.jl, a callback allows the ODE solver to detect an event and execute some action when that event occurs. This is useful for hybrid systems, switching circuits, threshold detection, impacts, resets, stopping conditions, etc. \[ \boxed{\text{condition} \longrightarrow \text{event} \longrightarrow \text{affect!}} \]

Callback Condition Typical use
ContinuousCallback (g(u,t)=0) zero crossings, thresholds, impacts
DiscreteCallback Boolean mode switching, logical conditions
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
function f!(du, u, p, t)
du[1] = -u[1]
end

u0 = [1.0]
tspan = (0.0, 10.0)

prob = ODEProblem(f!, u0, tspan)

function condition(u, t, integrator)
u[1] - 0.2
end

function affect!(integrator)
terminate!(integrator)
end

cb = ContinuousCallback(condition, affect!)

sol = solve(prob, Tsit5(), callback=cb)
println(sol.u[end][1]) # This will print the last value of u when the callback is triggered
println(sol.t[end]) # This will print the time at which the callback is triggered
println(exp(-sol.t[end])) # This will print the expected value of u at that time

Plots.plot(sol, xlims=(0, 2), linewidth=5, title="Solution of ODE with Callback", xlabel="Time", ylabel="u(t)")

# 0.20000000000000004
# 1.6094312935462547
# 0.20000132378195012

integrator is the currently running solver object. DifferentialEquations.jl automatically passes it into callback functions.

1
2
3
4
integrator.u  # current state
integrator.t # current time
integrator.p # problem parameters
integrator.sol # solution accumulated so far

ContinuousCallback

Use ContinuousCallback when the event is defined by a continuous zero crossing \(g(u,t)=0\)

graph LR
    A[ODE solver] --> B[integrate normally]
    B --> C{condition = 0 ?}
    C -- yes --> D["affect!()"]
    D --> E[continue integration]

image-20260815154158538 \[ \begin{cases} \dot{x} = v \\ \dot{v} = -g \end{cases} \]

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
using DifferentialEquations
import Plots


function ball!(du, u, p, t)
g = 9.81

du[1] = u[2] # dx/dt = v
du[2] = -g # dv/dt = -g
end

function condition(u, t, integrator)
u[1] # event when x = 0
end

function bounce!(integrator)
e = 0.8
integrator.u[2] = -e * integrator.u[2]
end

cb = ContinuousCallback(condition, bounce!)

u0 = [10.0, 0.0]

prob = ODEProblem(ball!, u0, (0.0, 10.0))

sol = solve(prob, Tsit5(), callback=cb)

Plots.plot(sol, size=(1200, 600), title="Bouncing Ball", xlabel="Time (s)", ylabel="Height (m)", legend=false)

DiscreteCallback

The condition returns a Boolean

1
condition(u, t, integrator) = true/false

DiscreteCallback checks its Boolean condition at the end of accepted integration steps. It does not use root finding to locate the exact point where \(u=1\)

image-20260815161659197

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
using DifferentialEquations
import Plots

function f!(du, u, p, t)
du[1] = 1.0
end

u0 = [0.0]
tspan = (0.0, 5.0)

prob = ODEProblem(f!, u0, tspan)

# Boolean condition
function condition_DT(u, t, integrator)
u[1] >= 1.0
end

function condition_CT(u, t, integrator)
u[1] - 1.0
end

# Action when condition == true
function affect!(integrator)
integrator.u[1] = 0.0
end

cb_DT = DiscreteCallback(condition_DT, affect!)
cb_CT = ContinuousCallback(condition_CT, affect!)

sol_DT = solve(prob, Tsit5(), callback=cb_DT)
println("Discrete-callback solution time: ", sol_DT.t[end-1:end])
println("Discrete-callback solution: ", sol_DT.u[end-1:end])
# Discrete-callback solution time: [5.0, 5.0]
# Discrete-callback solution: [[4.999999999999999], [0.0]]

sol_CT = solve(prob, Tsit5(), callback=cb_CT)

plt = Plots.plot(
sol_DT,
title="Two ODE Solutions: Discrete vs Continuous Callback",
xlabel="Time",
ylabel="u(t)",
label="Discrete-callback solution",
linewidth=2,
linestyle=:dash,
legend=:topleft,
size=(1200, 600),
)
Plots.plot!(plt, sol_CT, label="Continuous-callback solution", linewidth=2)

DiscreteCallback checks only after each accepted solver step. Because du/dt = 1 is exactly linear, Tsit5() takes a large step from approximately t=0.58 directly to t=5. It therefore does not check near u=1.


A callback can modify parameters

Callbacks provide the mechanism that connects the continuous ODE dynamics to this discrete switching behavior

So mathematically two components: \[ \dot{\mathbf{x}} = f(\mathbf{x}, p, t) \] for the \(\textbf{continuous-time dynamics}\), and \[ g(\mathbf{x}, t) = 0 \implies (\mathbf{x}, p) \to R(\mathbf{x}, p) \] for the \(\textbf{event/reset dynamics}\). \(R(x,p)\) means a reset map or event update rule

\[ \dot{x} = \begin{cases} -x, & x > 0.5 \\ -2x, & x < 0.5. \end{cases} \]

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
using DifferentialEquations
import Plots

function f!(du, u, p, t)
du[1] = -p[1]
end

function condition(u, t, integrator)
u[1] - 0.5
end

function affect!(integrator)
integrator.p[1] = -integrator.p[1]
end

p = [1.0]

cb = ContinuousCallback(condition, affect!)

prob = ODEProblem(f!, [1.0], (0.0, 1.0), p)

sol = solve(prob, Tsit5(), callback=cb)

Plots.plot(
sol,
title="A Callback Can Modify Parameters",
xlabel="Time (t)",
ylabel="State u₁(t)",
label="u₁(t)",
linewidth=2,
size=(1200, 600)
)

image-20260815164724091

reference

A. Demir, A. Mehrotra and J. Roychowdhury, "Phase noise in oscillators: a unifying theory and numerical methods for characterization," in IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications, vol. 47, no. 5, pp. 655-674, May 2000 [https://sci-hub.jp/10.1109/81.847872]

—, "A Reliable and Efficient Procedure for Oscillator PPV Computation, With Phase Noise Macromodeling Applications," IEEE TCAD, 2003.

— and A. Sangiovanni-Vincentelli, Analysis and Simulation of Noise in Nonlinear Electronic Circuits and Systems, vol. 425. Boston, MA, USA: Kluwer Academic Publishers, 1998

A. Mehrotra and A. Sangiovanni-Vincentelli, Noise Analysis of Radio Frequency Circuits, 1st ed. New York, NY, USA: Springer, 2004


Mathematical Preliminaries

Strogatz, S.H. (2015). Nonlinear Dynamics and Chaos: With Applications to Physics, Biology, Chemistry, and Engineering (2nd ed.). CRC Press [https://www.biodyn.ro/course/literatura/Nonlinear_Dynamics_and_Chaos_2018_Steven_H._Strogatz.pdf]

Higham, Desmond. (2001). An Algorithmic Introduction to Numerical Simulation of Stochastic Differential Equations. SIAM Review. 43. 525-546. 10.1137/S0036144500378302. [https://www.cmor-faculty.rice.edu/~cox/stoch/dhigham.pdf]

Jiří Lebl. Notes on Diffy Qs: Differential Equations for Engineers [link]

Matt Charnley. Differential Equations: An Introduction for Engineers [link]

Åström, K.J. & Murray, Richard. (2021). Feedback Systems: An Introduction for Scientists and Engineers Second Edition [https://www.cds.caltech.edu/~murray/books/AM08/pdf/fbs-public_24Jul2020.pdf]

Decimation Filter

[https://web.engr.oregonstate.edu/~temes/ece627/Lecture_Notes/First_Order_DS_ADC_scan1.pdf]

[https://web.engr.oregonstate.edu/~temes/ece627/Lecture_Notes/First_Order_DS_ADC_scan2.pdf]

The combination of the the digital post-filter and downsampler is called the decimation filter or decimator

image-20241015220921002

\(\text{sinc}\) filter

image-20241015215159577

Suppose \(T=1\) \[ H_1(e^{j2\pi f}) = \frac{\text{sinc}(Nf)}{\text{sinc}(f)} = \frac{1}{N}\frac{\sin(\pi Nf)}{\sin(\pi f)} \] that is \(\lim_{f\to 0^+}H_1(e^{j2\pi f}) = 1\) and \(H_1 = 0\) when \(f=\frac{n}{N}, n\in \mathbb{Z}\)

image-20241015215227042

A Beginner's Guide To Cascaded Integrator-Comb (CIC) Filters [https://www.dsprelated.com/showarticle/1337.php]

image-20241015225859710

image-20241015215111430

\[ |H_1(\omega)|^2 = \left|\frac{1}{N}(1-e^{-j\omega N}) \right| = \frac{2}{N^2}(1-\cos (\omega N)) \] Total noise after \(H_1\) \[ \sigma_{q_1}^2 = 2\int_0^\pi \frac{e^2_{rms}}{2\pi}\cdot |H_1(\omega)|^2d\omega = \frac{2e^2_{rms}}{N^2} \] inband noise before \(H_1\), i.e. ideal LPF with cutoff frequency \(\frac{\pi}{N}\) \[ \sigma_{q_0}^2 = 2\int_0^{\pi/N}\frac{e_{rms}^2}{2\pi}|1-e^{-j\omega}|^2d\omega = \frac{2e_{rms}^2}{\pi}\left(\frac{\pi}{N}-\sin\frac{\pi}{N}\right) \] with Taylor series \(\sin\frac{\pi}{N}\approx \frac{\pi}{N}-\frac{1}{6}\frac{\pi^3}{N^3}\) \[ \sigma_{q_0}^2 \approx \frac{\pi^2}{3N^3}e_{rms}^2 \]

Taylor’s Series of \(\sin x\) [pdf]

image-20250913093652192


[https://analogicus.com/aic2025/2025/02/20/Lecture-6-Oversampling-and-Sigma-Delta-ADCs.html#python-oversample]

\(\text{sinc}^2\) filter

image-20241015220030204

Interpolation Filter

Notice that the requirements of the first stage are very demanding

image-20250617001439043

replicas suppression

The spectrum of the high resolution digital signal \(u_1\) contains the original baseband portion and its replicas located at integer multiples of \(f_{s1}\), plus a small amount of quantization noise shown as a solid line

image-20250906170436567

image-20250918220425431


Nigel Redmon [https://dsp.stackexchange.com/a/63438/59253]

Inserting zeros changes nothing except what we consider the sample rate

image-20250920072501245

image-20250920074337023


Dan Boschen [https://dsp.stackexchange.com/a/32130/59253]

image-20250920075212621


Bourdopoulos, G. I. (2003). Delta-Sigma modulators : modeling, design and applications. Imperial College Press. [pdf]

image-20250920080910057

DC Gain in IF

DC gain is used to compensate the ratio of sampling rate before and after upsample

image-20250701070539064

Given \[ X_e = X = \propto \frac{1}{T} = \frac{1}{L\cdot T_i} \] Then, the lowpass filter (ZOH, FOH .etc) gain shall be \(L\)


Employ definition of DTFT, \(X(e^{j\hat{\omega}}) =\sum_{n=-\infty}^{+\infty}x[n]e^{-j\hat{\omega} n}\), and set \(\hat{\omega} = 0\) \[ X(e^{j0}) = \sum_{n=-\infty}^{+\infty}x[n] \] That is, \(\sum_{n=-\infty}^{+\infty}x[n] = \sum_{n=-\infty}^{+\infty}x_e[n]\), so \[ \overline{x_e[n]} = \frac{1}{L} \overline{x[n]} \] It also indicate that dc gain of upsampling is \(1/L\)

Zero-Order Hold (ZOH)

image-20250630235534325

dc gain = \(N\)

First-Order Hold (FOH)

image-20250630235714996

dc gain = \(N\)

Accumulate-and-dump (AAD) decimator

accumulating the input for \(N\) cycles and then latching the result and resetting the integrator

image-20241015222205883

It adds up \(N\) succeeding input samples at rate \(1/T\) and delivers their sum in a single sample at the output. Therefore, the process comprises a filter (in the accumulation) and a down-sampler (in the dump)

Cascaded Integrator-Comb (CIC) filter

Qasim Chaudhari, Cascaded Integrator Comb (CIC) Filters – A Staircase of DSP [https://wirelesspi.com/cascaded-integrator-comb-cic-filters-a-staircase-of-dsp/]

Tom Verbeure. An Intuitive Look at Moving Average and CIC Filters [https://tomverbeure.github.io/2020/09/30/Moving-Average-and-CIC-Filters.html]

—. Half-Band Filters, a Workhorse of Decimation Filters [https://tomverbeure.github.io/2020/12/15/Half-Band-Filters-A-Workhorse-of-Decimation-Filters.html]

—. Design of a Multi-Stage PDM to PCM Decimation Pipeline [https://tomverbeure.github.io/2020/12/20/Design-of-a-Multi-Stage-PDM-to-PCM-Decimation-Pipeline.html]

Arash Loloee, Ph.D. Exploring Decimation Filters [https://www.highfrequencyelectronics.com/Archives/Nov13/1311_HFE_decimationFilters.pdf]

Let's focus on decimation: if we decimate by a factor 4, we simply retain one output sample out of every 4 input samples.

In the example below, the downsampler at the right drops those 3 samples out of 4, and the output rate, \(y^\prime(n)\), is one fourth of the input rate \(x(n)\):

moving_average_filters-decimation_trivial \[\begin{align} Y(z) &= X(z)\frac{1-z^{-4}}{1-z^{-1}} \\ Y^\prime(\xi) &= \frac{1}{4}Y(\xi^{1/4}) = \frac{1}{4}X(\xi^{1/4})\frac{1-\xi^{-1}}{1-\xi^{-1/4}} \end{align}\]

with \(z=e^{j\Omega/f_s}\) and \(\xi =z^4\), we have \[ Y^\prime(z) = \frac{1}{4}X(z)\frac{1-z^{-4}}{1-z^{-1}} \]

But if we're going to be throwing away 75% of the calculated values, can't we just move the downsampler from the end of the pipeline to somewhere in the middle? Right between the integrator stage and the comb stage? That answer is yes, but to keep the math working, we also need to divide the number of delay elements in the comb stage by the decimation rate:

moving_average_filters-decimation_smart

\[\begin{align} A(z) &= X(z)\frac{1}{1-z^{-1}} \\ A^\prime(\xi) &= \frac{1}{4}A(\xi^{1/4}) = \frac{1}{4}X(\xi^{1/4})\frac{1}{1-\xi^{-1/4}} \\ Y^\prime(\xi) &= A^\prime(\xi) (1-\xi^{-1}) = \frac{1}{4}X(\xi^{1/4})\frac{1-\xi^{-1}}{1-\xi^{-1/4}} \end{align}\]

with \(z=e^{j\Omega/f_s}\) and \(\xi =z^4\), we have \[ Y^\prime(z) = \frac{1}{4}X(z)\frac{1-z^{-4}}{1-z^{-1}} \]


And we can do this just the same with cascaded sections (without downsampler or updampler) where integrators and combs have been grouped

  • for decimation, the integrators come first and the combs second with the downsampler in between
  • For interpolation, the reverse is true
    • the incoming sample rate is fraction of the outgoing sample rate, the combs must come first and the interpolators second

moving_average_filters-integrator_comb_decimated

moving_average_filters-comb_integrator_interpolated

reference

Pavan, Shanthi, Richard Schreier, and Gabor Temes. (2016) 2016. Understanding Delta-Sigma Data Converters. 2nd ed. Wiley.

K. Hosseini and M. P. Kennedy, Minimizing Spurious Tones in Digital Delta-Sigma Modulators (Analog Circuits and Signal Processing). New York, NY, USA: Springer, 2011.


Neil Robertson, Model a Sigma-Delta DAC Plus RC Filter [https://www.dsprelated.com/showarticle/1642.php]

—, Modeling a Continuous-Time System with Matlab [https://www.dsprelated.com/showarticle/1055.php]

—, Modeling Anti-Alias Filters [https://www.dsprelated.com/showarticle/1418.php]

—, DAC Zero-Order Hold Models [https://www.dsprelated.com/showarticle/1627.php]

—, “A Simplified Matlab Function for Power Spectral Density”, DSPRelated.com, March, 2020, [https://www.dsprelated.com/showarticle/1333.php]

Ahmed Shahein (2026). MSD Toolbox (https://github.com/ahmedshahein/MSDTOOLBOX), GitHub. Retrieved April 25, 2026.

—, Multi-Decimation Stage Filtering for Sigma Delta ADCs: Design and Optimization [https://www.dsprelated.com/showarticle/1037.php]

Rick Lyons. A Beginner's Guide To Cascaded Integrator-Comb (CIC) Filters [https://www.dsprelated.com/showarticle/1337.php]

—, Optimizing the Half-band Filters in Multistage Decimation and Interpolation [https://www.dsprelated.com/showarticle/903.php]


Venkatesh Srinivasan, ISSCC 2019 T5: Noise Shaping in Data Converters

Nan Sun,IEEE CAS 2020: Break the kT/C Noise Limit [https://www.facebook.com/ieeecas/videos/break-the-ktc-noise-limit/322899188976197/]

Yun-Shiang Shu, ISSCC 2022 T3: Noise-Shaping SAR ADCs

Xiyuan Tang, CICC 2025 ES2-1: Noise-Shaping SAR ADCs - From Fundamentals to Recent Advances

image-20250612003115259


Dual Slope ADC

image-20250615153045770

image-20250615152228233

\[ V_{IN} = \frac{V_{REF}}{T}t_\text{x} = \frac{V_{REF}}{2^N}\cdot 2^{N_\text{x}} \]

Normal Mode Rejection

a high normal mode rejection ratio (NMRR) for input noise at line frequency

image-20250615160802268

  • Conversion accuracy is independent of both the capacitance and the clock frequency, because they affect both the up-slope and the down-slope by the same ratio

  • The fixed input signal integration period results in rejection of noise frequencies on the analog input that have periods that are equal to or a sub-multiple of the integration time \(T\)

    Interference signals with frequencies at integral multiples of the integration period are, theoretically, completely removed, since the average value of a sine wave of frequency (\(1/T\)) averaged over a period (\(T\)) is zero

image-20250615155921455

Linear Circuit Design Handbook, 2008 [https://www.analog.com/media/en/training-seminars/design-handbooks/Basic-Linear-Design/Chapter6.pdf]

Precision Analog Front Ends with Dual Slope ADC [https://ww1.microchip.com/downloads/aemDocuments/documents/APID/ProductDocuments/DataSheets/21428e.pdf]

Incremental ADC

Z. Tan, C. -H. Chen, Y. Chae and G. C. Temes, "Incremental Delta-Sigma ADCs: A Tutorial Review," in IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 67, no. 12, pp. 4161-4173, Dec. 2020 [https://sci-hub.jp/10.1109/TCSI.2020.3033458]

image-20250615124340044

image-20250615162024564

image-20250615162339449

image-20250615162358247

image-20250615170209562

\[\begin{align} V &= 2^N V_\text{in} - D_\text{out}V_\text{ref} \\ D_\text{out} \frac{V_\text{ref}}{2^N} &= V_\text{in} - \frac{V}{2^N} \end{align}\]

image-20250615164436626

image-20250615192031025

feedforward structure

??? TODO 📅

image-20250615165107485

image-20250615175452958


image-20250615194549768

image-20250615194825368

image-20250615195150811

reference

David Johns (University of Toronto) "Oversampled Data Converters" Course (2019) [https://youtu.be/qIJ2LORYmyA]

Maurits Ortmanns , Paul Kaesser , Johannes Wagner (Dec 2025). Incremental Delta-Sigma ADCs Theory, Architectures and Design - Theory, Architectures and Design

image-20260914193314752

Frequency Domain Model

f-mdl.drawio

image-20260913094812184

Why (8.14) means "sampled at \(T_s\)" \[ X_s(f) = \frac{1}{T_s}\sum_{k=-\infty}^{\infty} X(f - kf_s) \]

This spectrum is periodic with period \(f_s\): copies of \(X(f)\) sit at every multiple of \(f_s\), with weight \(1/T_s\). That is exactly the spectrum of uniform sampling at \(T_s\)

Interleaver Architectures

image-20260914194340521

Direct Interleaver

image-20260914194944454

similar to increase the resolution of the flash ADC with more parallel comparators

De-multiplexing Interleaver

image-20260914200548166

it is the front-end samplers that determine timing/bandwidth mismatch errors

only one front-end channel \(L=1\) eliminate any timing/bandwidth mismatch errors to the first order

Re-sampling Interleaver

image-20260914195228805

back-end re-sampling occur after the front-end, two \(\frac{KT}{C}\) contribution in total noise (De-multiplexing Interleaver only one \(\frac{KT}{C}\))

without buffer, charging distribution reduce signal and reduce SNR, but buffers give excess noise


image-20260906160455591

Interleaver Model

image-20260914195326519

Interleaving Errors

image-20260914202358169

image-20260915213004819

Offset Mismatch Errors

image-20260914202446007

Gain Mismatch Errors

image-20260914202550362

image-20250621072944516

image-20260914210022880

Timing Mismatch Errors

image-20260914204512858

In the frequency domain,

\[ \mathcal{L}\left\{ \frac{dV(t)}{dt} \right\} \;\longrightarrow\; sV(s) \quad\text{and, for }s=j\omega,\quad j\omega V(j\omega). \]

The factor

\[ j=e^{j\pi/2} \]

introduces a \(+\pi/2\) (90°) phase shift, while \(\omega\) scales the magnitude proportionally to frequency.

In the time domain,

\[ \frac{d}{dt}\sin(\omega t) =\omega\cos(\omega t) =\omega\sin\left(\omega t+\frac{\pi}{2}\right). \]

Thus, differentiation produces two effects: magnitude scaling by \(\omega\) and a \(90^\circ\) phase advance.

Frequency-dependent: the higher frequency input signal \(f_\text{in}\), the larger error becomes

image-20250621091024424

image-20260914210257078

image-20250621091047339

Bandwidth Mismatch Errors

image-20260914202752033

image-20250621093321623

ADC Clocking

Track-and-Hold (TAH), Sample-and-Holds (SAH)

image-20261017111944790

Note that the falling edge of \(\phi_1\) is determined by the rising edge of CK and is thus free from jitter and mismatch in \(X_1\), which accumulates jitter and phase mismatch

image-20260920233403464

reset mechanism to the latches for nominal order

image-20261017112345989


a clock generator for an eight-channel ADC

image-20260920235134383

NORed \(\div 2\) \(\div 2 \div 2\)
\(\phi_1\) \(\text{CK}\) \(I_1\) \(I_2\)
\(\phi_2\) \(\overline{\text{CK}}\) \(Q_1\) \(I_3\)
\(\phi_3\) \(\text{CK}\) \(\overline{I_1}\) \(Q_2\)
\(\phi_4\) \(\overline{\text{CK}}\) \(\overline{Q_1}\) \(Q_3\)
\(\phi_5\) \(\text{CK}\) \(I_1\) \(\overline{I_2}\)
\(\phi_6\) \(\overline{\text{CK}}\) \(Q_1\) \(\overline{I_3}\)
\(\phi_7\) \(\text{CK}\) \(\overline{I_1}\) \(\overline{Q_2}\)
\(\phi_8\) \(\overline{\text{CK}}\) \(\overline{Q_1}\) \(\overline{Q_3}\)

\(\text{CK}\), \(I/Q_1\) and \(I/Q_{2,3}\) shift by 1UI

Interchannel Crosstalk

D. Xu et al., "8.5 A Scalable Adaptive ADC/DSP-Based 1.25-to-56Gbps/112Gbps High-Speed Transceiver Architecture Using Decision-Directed MMSE CDR in 16nm and 7nm," 2021 IEEE International Solid-State Circuits Conference (ISSCC), San Francisco, CA, USA, 2021, pp. 134-136 [https://sci-hub.ru/10.1109/ISSCC42613.2021.9366063]

image-20260920232810972

image-20261018192820211

64Gb/s Clock Generation

D. Xu et al., "8.5 A Scalable Adaptive ADC/DSP-Based 1.25-to-56Gbps/112Gbps High-Speed Transceiver Architecture Using Decision-Directed MMSE CDR in 16nm and 7nm," 2021 IEEE International Solid-State Circuits Conference (ISSCC), San Francisco, CA, USA, 2021, pp. 134-136 [https://sci-hub.ru/10.1109/ISSCC42613.2021.9366063]

image-20261018193058479

image-20261018192820211


4x8 Time Interleaving w/ 25% duty cycle for input tracking

adc_clocking_TH

4way-clocking-TH.drawio

Ensure nominal order by masking some edges

adc_clocking_reset

4way-clocking-reset-mask.drawio

128Gb/s Clock Generation

Z. Guo et al., "A 112.5Gb/s ADC-DSP-Based PAM-4 Long-Reach Transceiver with >50dB Channel Loss in 5nm FinFET," 2022 IEEE International Solid-State Circuits Conference (ISSCC), San Francisco, CA, USA, 2022 [https://sci-hub.st/10.1109/ISSCC42614.2022.9731650]

P. Liu et al., "A 128Gb/s ADC/DAC Based PAM-4 Transceiver with >45dB Reach in 3nm FinFET," 2025 Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits), Kyoto, Japan, 2025

adc-clk_4to32way_8to64way.drawio

image-20250806224145281

8way-interleaving-Marvell-ISSCC2022.drawio

image-20250808220954208

image-20250808221134749

RX-Clocking-Marvell-VLSI2025.drawio

224Gb/s Clock Generation

M. Gambhir et al., "A 2.05pJ/bit, 212.5Gbps DSP Based Transceiver with 55dB Reach in 3nm FinFET," 2026 IEEE/JSAP Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits), Honolulu, HI, USA, 2026, pp. 1-3, doi: 10.1109/VLSITechnologyandCir65830.2026.11577411.

image-20261018203306919

image-20261018203332571

4UI track, 12UI hold in TAH

shift register-based clock divider for TH2 clocks

J. Im et al., "6.1 A 112Gb/s PAM-4 Long-Reach Wireline Transceiver Using a 36-Way Time-Interleaved SAR-ADC and Inverter-Based RX Analog Front-End in 7nm FinFET," 2020 IEEE International Solid-State Circuits Conference - (ISSCC), San Francisco, CA, USA, 2020, pp. 116-118 [https://sci-hub.ru/10.1109/ISSCC19947.2020.9063081]

—, "A 112-Gb/s PAM-4 Long-Reach Wireline Transceiver Using a 36-Way Time-Interleaved SAR ADC and Inverter-Based RX Analog Front-End in 7-nm FinFET," in IEEE Journal of Solid-State Circuits, vol. 56, no. 1, pp. 7-18, Jan. 2021 [https://sci-hub.ru/10.1109/JSSC.2020.3024261]

image-20261025125422783

image-20261031123020616

thck_6ui_2p4ui.drawio

image-20261025130744430

image-20261025135125595

th2ck_36ui_4ui.drawio

image-20261031123046608

image-20261025143050494


a.k.a. ring counter, generating SAH clock

image-20261018235719799

adc_clocking_ring

all FFs initialized to 0

Rank1 & Rank2 Clock relationship

J. Im et al., "6.1 A 112Gb/s PAM-4 Long-Reach Wireline Transceiver Using a 36-Way Time-Interleaved SAR-ADC and Inverter-Based RX Analog Front-End in 7nm FinFET," 2020 IEEE International Solid-State Circuits Conference - (ISSCC), San Francisco, CA, USA, 2020, pp. 116-118 [https://sci-hub.ru/10.1109/ISSCC19947.2020.9063081]

—, "A 112-Gb/s PAM-4 Long-Reach Wireline Transceiver Using a 36-Way Time-Interleaved SAR ADC and Inverter-Based RX Analog Front-End in 7-nm FinFET," in IEEE Journal of Solid-State Circuits, vol. 56, no. 1, pp. 7-18, Jan. 2021 [https://sci-hub.ru/10.1109/JSSC.2020.3024261]

M. -A. LaCroix et al., "8.4 A 116Gb/s DSP-Based Wireline Transceiver in 7nm CMOS Achieving 6pJ/b at 45dB Loss in PAM-4/Duo-PAM-4 and 52dB in PAM-2," 2021 IEEE International Solid-State Circuits Conference (ISSCC), San Francisco, CA, USA, 2021, pp. 132-134 [https://sci-hub.ru/10.1109/ISSCC42613.2021.9366030]

G. Li et al., "18.1 A 600Gb/s DP-QAM64 Coherent Optical Transceiver Frontend with 4x105GS/s 8b ADC/DAC in 16nm CMOS," 2024 IEEE International Solid-State Circuits Conference (ISSCC), San Francisco, CA, USA, 2024, pp. 338-340 [https://sci-net.xyz/10.1109/ISSCC49657.2024.10454499]

image-20261031125216446

There are overlap between clk6UI_ph0UI and clk36UI_ph0UI, clk36UI_ph12UI, clk36UI_ph24UI, clk36UI_ph6UI, clk36UI_ph18UI, clk36UI_ph30UI

J. Im: TH2 track overlaps TH1 track

56GS/s, 1UI = 17.9ps window
TH1 track (40% DCC) \(0 \to 2.4\text{UI}\)
TH1 hold \(2.4 \to 6\text{UI}\) (3.6UI)
TH2 track (ring counter by ph0, trim by ph4) \(0 \to 4\text{UI}\)
  • overlap with TH1 current track: 2.4UI
  • TH2 settling after TH1 hold: \(4-2.4 = 1.6\text{UI}\) (~29ps)
  • margin before TH1 next track: \(6-4 = 2\text{UI}\) (~36ps)

what matters is where \(\Phi_2\) falling edge lands: after \(\Phi_1\) falling edge + settling, before \(\Phi_1\) next rising edge with margin for PVT drift between fast Rank1 path & slow Rank2 path → LaCroix's aligner

image-20261031125418057

image-20261031125520312

ADC Data Alignment

Gain Kim, 2023. Equalization, Architecture, and Circuit Design for High-Speed Serial Link Receiver [www.theise.org/...]

image-20261009070618262

Skew Calibration Techniques

Timing-skew calibration has two separate tasks: detect the relative sampling-time errors, then correct the sampling instants or the resulting digital errors. Foreground/background describes when estimation runs; analog/digital describes where correction occurs. A digital detector driving a variable delay line is a mixed-signal calibration loop.

For an \(M\)-channel TI-ADC, let \(f_s=1/T_s\) be the aggregate sampling rate and \(\tau_i\) the skew of channel \(i\):

\[ t_{i,k}=(kM+i)T_s,\qquad y_i[k]=x(t_{i,k}+\tau_i) \approx x(t_{i,k})+\tau_i\dot{x}(t_{i,k}). \]

This first-order model assumes small skew and omits other errors. It explains why timing error increases with input slope. Calibration removes relative skew: adding the same delay to every channel shifts the time origin without producing interleaving mismatch. Random aperture jitter and frequency-dependent bandwidth mismatch are separate limitations. See S1 and S2.

Method Comparison

The main background families are input-statistics-based, reference-channel-based, and reference-signal-based [S1]. The rows below separate useful implementations within these families; digital correction is discussed separately because it can accompany several detectors.

Method Timing-error observable Main advantage Main limitation / cost Read
Foreground training Channel phase or sample error for a known sine, ramp, or reference edge Simple identification and startup calibration Interrupts normal conversion; stored trims do not track later drift S1, S2
Autocorrelation / MAD Unequal statistics of nominally equal sample intervals No extra analog reference channel Input statistics, coherent patterns, and detector slope affect convergence S3, S4
Digital mixing Products of channel outputs and an estimate of the autocorrelation slope Can estimate skew magnitude for digital correction Derivative estimation and correction bandwidth; some implementations add auxiliary samplers S5, S6
Zero-crossing statistics Unequal crossing counts between adjacent sampling instants Simple sign/count logic Requires crossing activity and appropriate sample-sequence control S7
Reference-ADC correlation / variance Correlation or difference between nominally coincident main/reference samples Provides a common timing reference for all channels Extra input loading, reference-clock quality, averaging, and possible reference spurs S8, S9
Reference-ADC derivative / LMS Main/reference sample error multiplied by input slope Direct early/late information Must obtain a useful slope estimate; weak input activity slows adaptation S10
Injected reference / dither Response to a known timing-sensitive signal correlated with its modulation sequence Supplies excitation independently of the unknown input Injection-path mismatch, headroom, kickback, and residual dither S11, S12

Foreground Calibration with a Known Input

Apply a known calibration waveform, estimate each channel's relative timing, and store clock-trim codes or digital correction coefficients. For a sinusoid, an unwrapped phase difference gives \(\Delta\tau=\Delta\phi/(2\pi f_{\mathrm{cal}})\) after accounting for the nominal interleaving phase. A ramp converts timing error into a voltage error through its known slope.

This is useful for initial characterization and coarse startup alignment. Offset, gain, and analog-path phase must be separated from timing error; otherwise the trim can absorb another mismatch. Recalibration is needed if the relative delays drift. Foreground coefficients may be applied continuously during operation, but that alone does not make their estimation a background process [S1], [S2].

Autocorrelation-Based Skew Calibration

Assume the input is wide-sense stationary over the averaging interval, with

\[ R_x(\tau)=E[x(t)x(t+\tau)]. \]

After offset/gain equalization, the correlation between adjacent channel samples is approximately

\[ C_i=R_x(T_s+\tau_{i+1}-\tau_i) \approx R_x(T_s)+R_x'(T_s)(\tau_{i+1}-\tau_i). \]

For the last channel, use channel 0's sample from the next interleaving frame and set \(\tau_M=\tau_0\). Equal nominal intervals should give equal correlations; deviations reveal relative skew when \(R_x'(T_s)\) is observable.

Hierarchical Calibration and Its Bandwidth Limit

A hierarchical scheme first aligns widely separated channels, then fills in the intermediate phases. Its detector can therefore depend on \(R_x'(L T_s)\) with \(L>1\), whose sign is harder to guarantee over a wide input band. In the power-of-two hierarchy discussed in [S3], the first step uses a lag of \(M T_s/2\); its usual input-bandwidth guarantee is limited to \(f_s/M\) [S1].

Channel grouping and trim-code search are different choices. Binary search over delay codes still needs a reliable, monotonic error indication; it does not remove a detector's input-bandwidth restriction.

image-20260328225656118

Adjacent-Channel Calibration

Gu et al. [S4] use adjacent-channel statistics to extract all channel skews without the long-lag first step. The first-order relationship can be written as

\[ \overline{C}=\frac{1}{M}\sum_{i=0}^{M-1}C_i\approx R_x(T_s), \qquad \sum_{i=0}^{m-1}(C_i-\overline{C}) \approx R_x'(T_s)(\tau_m-\tau_0). \]

The sum telescopes, giving the skew of channel \(m\) relative to channel 0. This illustrates why adjacent correlations can support simultaneous estimation and extend the applicable bandwidth toward \(f_s/2\).

image-20260330204902627

image-20260330205119231

Nyquist-band coverage is not input independence. The ensemble model and finite per-channel averages can differ for coherent inputs. The review [S1] explicitly identifies a sinusoid at \(f_s/M\) as a problematic case for [S4]. Low signal activity also reduces detector gain.

For a random-phase sinusoid of amplitude \(A\),

\[ R_x(\tau)=\frac{A^2}{2}\cos(2\pi f_{\mathrm{in}}\tau), \qquad R_x'(T_s)=-\pi A^2 f_{\mathrm{in}}\sin(2\pi f_{\mathrm{in}}T_s). \]

Thus the ideal adjacent-correlation detector loses first-order sensitivity at DC and at an exact Nyquist tone. A result measured near Nyquist does not establish convergence for every coherent pattern.

image-20260324183003185


Why the Wei–Razavi method has a bandwidth restriction

Using the correlation model discussed above, a detector comparing intervals \(L+\Delta t\) and \(L-\Delta t\) produces

\[ e=R_x(L+\Delta t)-R_x(L-\Delta t) \approx 2\Delta t\,R_x'(L) \]

Therefore, the sign of \(e\) indicates the sign of skew only if the sign of \(R_x'(L)\) is known.

For a real WSS input with two-sided PSD \(S_x(f)\) confined to \(|f|\le B\),

\[ R_x'(L) =-4\pi\int_0^B fS_x(f)\sin(2\pi fL)\,df \]

A sufficient condition for a negative slope, assuming nonzero signal energy inside the band, is

\[ \boxed{B<\frac{1}{2L}} \]

Above that range, positive and negative contributions can cancel or reverse the slope.

In the hierarchical multichannel extension, the first calibration step uses \(L=MT_s/2\), giving the familiar guarantee

\[ \boxed{B<\frac{f_s}{M}} \]

The Gu's 2025 review explicitly identifies this limitation. It is a detector-polarity guarantee, rather than a claim that every higher-frequency input necessarily fails

Gu: shorten the correlation lag to address the restriction

Approach What changes Remaining limitation
Adjacent-channel correlations — Gu et al., JSSC 2025 [S4] Uses only lag \(T_s\), extending the correlation-slope bandwidth condition toward \(f_s/2\) Still depends on input statistics; certain coherent tones remain problematic

The method measures all adjacent-channel correlations, uses their average to estimate the ideal adjacent correlation, and combines the deviations to recover each channel’s relative skew. It avoids the widely separated channel pairs in the first hierarchical step:

\[ \boxed{ L:\frac{MT_s}{2}\rightarrow T_s \quad\Longrightarrow\quad B_{\mathrm{guarantee}}:\frac{f_s}{M}\rightarrow\frac{f_s}{2} } \]

However, the review still identifies a coherent sinusoid at \(f_s/M\) as a failure case. Extending the bandwidth does not eliminate sampling-pattern dependence.

MAD (Mean Absolute Difference) vs. Correlation

Wei et al. [S3] compare the mean absolute differences on the two sides of a sample, avoiding general-purpose multipliers. To understand the relationship, define

\[ D(\tau)=E[|x(t+\tau)-x(t)|],\qquad Q(\tau)=E[(x(t+\tau)-x(t))^2]. \]

For a real WSS signal, the squared-difference identity is exact:

\[ Q(\tau)=2[R_x(0)-R_x(\tau)]. \]

MAD does not equal the squared difference. Replacing the absolute value by a square in [S3] is an analytical surrogate. If the increment is zero-mean Gaussian, then \(D(\tau)=\sqrt{2Q(\tau)/\pi}\), so the two metrics are monotonic functions of one another. WSS alone does not guarantee that relation for arbitrary input distributions.

image-20260328212413965

Digital-Mixing-Based Calibration

Here, digital mixing forms products of channel samples to measure correlation imbalance. It belongs to the statistical family above. To estimate skew magnitude, rather than only an update direction, the imbalance must be normalized by an estimate of the relevant correlation slope.

Lin et al. [S5] obtain the required information with auxiliary delta-sampling SAR ADCs. Guo et al. [S6] estimate the autocorrelation derivative digitally and correct the timing error without those extra analog channels. Their first-order approximation and derivative bandwidth deserve attention, especially for direct sampling above the first Nyquist zone. The reported 29 mW in [S6] is ADC-core power with mismatch calibration implemented off-chip, not the power of a complete on-chip calibration system.

The existing notes from [S6] are retained here:

image-20260324180833001

Zero-Crossing-Based Calibration

Wang and Wu [S7] detect skew by counting input zero crossings between adjacent channels while randomly alternating their sampling order. The count imbalance drives digitally controlled delays. This uses sign decisions and counters instead of full-resolution products.

The input must provide sufficient crossings, and the sampling-order scheme is part of the detector. A generic crossing counter cannot be assumed to recover skew for every waveform or sampling rate. Study the paper's convergence and residual-jitter analysis together with its switching sequence.

Reference-Channel-Based Calibration

An extra ADC samples the same input at nominally coincident instants with each main channel. For an already gain/offset-matched pair, a local model is

\[ e_i=y_i-y_{\mathrm{ref}} \approx (\tau_i-\tau_{\mathrm{ref}})\dot{x}. \]

The reference may run slowly while visiting all channel phases. In [S8], choosing \(f_{\mathrm{CAL}}=f_s/K\) with \(\gcd(K,M)=1\) provides that coverage. Coprime division guarantees phase-slot coverage, not automatic physical edge alignment; the clock origin and correction range still need to be designed.

Correlation Maximization / Difference-Variance Minimization

El-Chammas and Murmann [S8] maximize the cross-correlation between the reference and each main channel. With matched scaling and delay-independent signal power, maximizing correlation is equivalent to minimizing mean-squared sample difference. Song et al. [S9] provide a variance-based implementation to study alongside it.

These metrics have an extremum at alignment, so a search or perturbation is needed to determine which direction to move. Reference-path bandwidth mismatch can move the optimum; extra sampling loads can also disturb the main input.

Derivative-Based LMS

Multiplying the error by an input-slope estimate gives a signed timing observable:

\[ E[e_i\dot{x}] \approx (\tau_i-\tau_{\mathrm{ref}})E[\dot{x}^{\,2}]. \]

A negative-feedback update can therefore drive the relative skew toward zero. Its actual code-update sign depends on whether increasing the delay code advances or retards the sampling edge.

Stepanović and Nikolić [S10] estimate the derivative using two calibration paths with different analog bandwidths. Read this for the implementation of the slope detector and LMS loop. Unlike purely statistical correlation search, the slope provides direct directional information, but adds circuitry and still needs input activity.

Reference-Signal / Dither-Injection Calibration

Inject a known waveform with a timing-sensitive slope into the input path. In foreground operation it can replace the normal input. For background operation, modulate the reference with a known pseudo-random sequence and correlate the ADC outputs with that sequence to extract its contribution [S11], [S12].

This supplies calibration excitation even when the external input is quiet or unfavorable for a statistical detector. The main design issue is which path the injected signal measures. Injection after a mismatched input-buffer stage can miss that stage's delay; separate injection paths can introduce their own mismatch. Tao et al. [S11] and Cao et al. [S12] are useful comparisons of shared/global injection arrangements.

Check the residual injected component, input headroom, reference-edge quality, and disturbance of the wanted signal. These determine the cost of improved input independence.

Clock Trimming vs. Digital Error Correction

Correction Implementation What to check
Analog sampling-time correction Digital detector adjusts a variable delay line, phase interpolator, or clock-buffer trim Delay range, resolution, monotonicity, added jitter, and whether the adjusted edge controls the actual sampling aperture
Digital reconstruction Derivative cancellation or fractional-delay / reconstruction filters correct the output samples Accurate skew magnitude, filter bandwidth, higher-order residuals, word length, throughput, and power

For small skew, the first-order digital correction is

\[ \widehat{x}[n]=y[n]-\widehat{\tau}_{i(n)}\,\widehat{\dot{x}}[n], \qquad i(n)=n\bmod M, \]

where \(\widehat{\dot{x}}\) estimates the derivative with respect to physical time. A derivative per sample instead requires the normalized coefficient \(\widehat{\tau}_i/T_s\). Align the derivative-filter latency with the data and channel index. The remaining error depends on derivative accuracy and higher-order Taylor terms; a digital detector alone does not guarantee an all-digital solution. Read S13, S6, and the earlier adaptive approach in S14.

A common front-end track-and-hold can reduce downstream aperture-mismatch sensitivity, while channel randomization can spread mismatch spurs. These are architectural mitigation techniques, not estimates of the skew itself; spur spreading does not by itself remove total error power [S1].

For an ADC-based SerDes, a useful design check is to identify which front-end sampling rank sets the input aperture, then evaluate the detector with the actual equalized data statistics and repeating patterns. Common sampling-phase recovery and relative interleaver-skew calibration control different degrees of freedom.

Papers to Read

Suggested first pass: [S1] → [S2] → [S3] → [S4] for statistical detection, then [S8] → [S10] → [S11] to compare reference-assisted approaches. Read [S13] → [S6] for digital correction. The remaining papers deepen particular branches.

[S1] Overview. M. Gu, Y. Tao, Y. Zhong, L. Jie, and N. Sun, “Timing-Skew Calibration Techniques in Time-Interleaved ADCs,” IEEE Open Journal of the Solid-State Circuits Society, vol. 5, pp. 1–10, 2025. DOI / open-access paper. Read first for the taxonomy and input-dependence tradeoffs. Note that this review's author list differs from [S4].

[S2] Fundamentals. B. Razavi, “Problem of Timing Mismatch in Interleaved ADCs,” IEEE CICC, 2012. Author PDF. Read for the error model, sources of skew, and detection/correction alternatives.

[S3] MAD detector and delay correction. H. Wei, P. Zhang, B. D. Sahoo, and B. Razavi, “An 8 Bit 4 GS/s 120 mW CMOS ADC,” IEEE JSSC, vol. 49, no. 8, pp. 1751–1761, Aug. 2014. DOI, author PDF. The earlier CICC 2013 paper 16.3 is here. Focus on the multiplier-free detector, channel hierarchy, and fine delay line.

[S4] Adjacent-channel statistics. M. Gu, Y. Tao, X. He, Y. Zhong, L. Jie, and N. Sun, “A 1-GS/s 11-b Time-Interleaved SAR ADC With Robust, Fast, and Accurate Autocorrelation-Based Background Timing-Skew Calibration,” IEEE JSSC, vol. 60, no. 2, pp. 421–431, Feb. 2025. DOI. Compare its simultaneous extraction and bandwidth with [S3], and examine the remaining input restrictions.

[S5] Digital mixing with auxiliary sampling. C.-Y. Lin, Y.-H. Wei, and T.-C. Lee, “A 10-bit 2.6-GS/s Time-Interleaved SAR ADC With a Digital-Mixing Timing-Skew Calibration Technique,” IEEE JSSC, vol. 53, no. 5, pp. 1508–1517, May 2018. DOI. Focus on the delta-sampling auxiliary ADCs and how skew magnitude is obtained.

[S6] Fully digital mixing-based calibration. M. Guo, J. Mao, S.-W. Sin, H. Wei, and R. P. Martins, “A 5 GS/s 29 mW Interleaved SAR ADC With 48.5 dB SNDR Using Digital-Mixing Background Timing-Skew Calibration for Direct Sampling Applications,” IEEE Access, vol. 8, pp. 138944–138954, 2020. DOI / open-access paper. Focus on the autocorrelation-derivative approximation and off-chip digital correction.

[S7] Zero-crossing statistics. C.-Y. Wang and J.-T. Wu, “A Background Timing-Skew Calibration Technique for Time-Interleaved Analog-to-Digital Converters,” IEEE TCAS-II: Express Briefs, vol. 53, no. 4, pp. 299–303, Apr. 2006. DOI, institutional PDF. Read for sampling-sequence randomization, crossing counts, and loop convergence.

[S8] Reference-ADC correlation. M. El-Chammas and B. Murmann, “A 12-GS/s 81-mW 5-bit Time-Interleaved Flash ADC With Background Timing Skew Calibration,” IEEE JSSC, vol. 46, no. 4, pp. 838–847, Apr. 2011. DOI. Read together with B. Murmann and M. El-Chammas, “Background Calibration of Timing Skew in Time-Interleaved A/D Converters,” SampTA, 2011 (author PDF), which explains the coprime reference-clock schedule and correlation objective.

[S9] Variance-based reference calibration. J. Song, K. Ragab, X. Tang, and N. Sun, “A 10-b 800-MS/s Time-Interleaved SAR ADC With Fast Variance-Based Timing-Skew Calibration,” IEEE JSSC, vol. 52, no. 10, pp. 2563–2575, Oct. 2017. DOI, author page. Focus on its single-comparator window detector and faster variance estimation; compare the objective with [S8].

[S10] Analog derivative and LMS. D. Stepanović and B. Nikolić, “A 2.8 GS/s 44.6 mW Time-Interleaved ADC Achieving 50.9 dB SNDR and 3 dB Effective Resolution Bandwidth of 1.5 GHz in 65 nm CMOS,” IEEE JSSC, vol. 48, no. 4, pp. 971–982, Apr. 2013. DOI, author PDF. Focus on the two calibration paths, derivative estimate, and timing-control loop.

[S11] Background reference injection. Y. Tao, M. Gu, B. Chi, Y. Zhong, L. Jie, and N. Sun, “22.4 A 4.8GS/s 7-ENoB Time-Interleaved SAR ADC with Dither-Based Background Timing-Skew Calibration and Bit-Distribution-Based Background Ping-Pong Comparator Offset Calibration,” IEEE ISSCC, pp. 394–396, 2024. DOI. Focus on reference modulation, injection location, and cancellation of the injected component.

[S12] Global dither injection. Y. Cao, M. Zhang, Y. Zhu, R. P. Martins, and C.-H. Chan, “A 12-GS/s 12-b 4× Time-Interleaved ADC Using Input-Independent Timing Skew Calibration With Global Dither Injection and Linearized Input Buffer,” IEEE JSSC, vol. 59, no. 12, pp. 4211–4224, Dec. 2024. DOI. Read for the difference between local and global injection and the timing errors each can observe.

[S13] All-digital estimation and correction. S. Chen, L. Wang, H. Zhang, R. Murugesu, D. Dunwell, and A. Chan Carusone, “All-Digital Calibration of Timing Mismatch Error in Time-Interleaved Analog-to-Digital Converters,” IEEE Transactions on VLSI Systems, vol. 25, no. 9, pp. 2552–2560, Sept. 2017. DOI, author PDF, slides. Focus on difference-based estimation, derivative-filter correction, and residual SNDR limits.

[S14] Early adaptive digital calibration. S. M. Jamal, D. Fu, N. C.-J. Chang, P. J. Hurst, and S. H. Lewis, “A 10-b 120-Msample/s Time-Interleaved Analog-to-Digital Converter With Digital Background Calibration,” IEEE JSSC, vol. 37, no. 12, pp. 1618–1627, Dec. 2002. DOI. Read for correlation-driven adaptive correction of gain and sample-time errors.

Overlapping versus Non-overlapping track time

image-20250611224418950

tracking accuracy stay same, Cin (2Cs) counteract the longer tracking

Summing Interleaved Alias

image-20240929215841300

The sampling function - impulse train is \[ s(t) = \sum_{n=-\infty}^{\infty}\left[ \delta(t-n4T_s) + \delta(t-n4T_s-T_s) + \delta(t-n4T_s-2T_s) + \delta(t-n4T_s-3T_s)\right] \]

Its Fourier transform is \[\begin{align} S(f) &= \frac{2\pi}{4T}\sum_{k=-\infty}^{\infty}\left[\delta(f-k\frac{f_s}{4}) + e^{-j2\pi f\cdot T_s}\delta(f-k\frac{f_s}{4}) + e^{-j2\pi f\cdot 2T_s}\delta(f-k\frac{f_s}{4}) + e^{-j2\pi f\cdot 3T_s}\delta(f-k\frac{f_s}{4}) \right] \\ &= \frac{2\pi}{4T}\sum_{k=-\infty}^{\infty}\left(1+e^{-j2\pi\frac{f}{f_s}} + e^{-j4\pi\frac{f}{f_s}} + e^{-j6\pi\frac{f}{f_s}} \right) \delta(f-k\frac{f_s}{4}) \\ &= \frac{2\pi}{4T}\sum_{k=-\infty}^{\infty}\left(1+e^{-jk\frac{\pi}{2}} + e^{-jk\pi} + e^{-jk\frac{3\pi}{2}} \right) \delta(f-k\frac{f_s}{4}) \end{align}\]

We define \(M[k] = 1+e^{-jk\frac{\pi}{2}} + e^{-jk\pi} + e^{-jk\frac{3\pi}{2}}\), which is periodic, i.e. \(M[k]=M[k+4]\) \[ M[k]=\left\{ \begin{array}{cl} 4 & : \ k = 4m \\ 0 & : \ k=4m+1 \\ 0 & : \ k=4m+2 \\ 0 & : \ k=4m+3 \\ \end{array} \right. \]

That is \[ S(f) = \frac{2\pi}{T}\sum_{k=-\infty}^{\infty} \delta(f-kf_s) \]

Alias has same frequency for each slice but different phase: Alias terms sum to zero if all slices match exactly

Random Chopping in TI-ADC

image-20240929215927957

\[ D_n(kT) = (G_n R(kT) V(kT) + O_n)R(kT)= C_n V(kT) + R(kT)O_n \]

DNL/INL Benefit

Each sub-ADC has its own physical capacitor array, so each slice has its own independent, zero-mean mismatch error

Because the sub-ADC are sampled round-robin, a busy input signal visits all \(N\) sub-ADC with equal probability over the whole input range. The transfer curve that the composite output presents is therefore the ensemble average \[ e_{\mathrm{eff}}(k) = \frac{1}{N}\sum_{i=1}^{N} e_i(k) \quad\Rightarrow\quad \sigma\!\left\{e_{\mathrm{eff}}(k)\right\} = \frac{\sigma\{e_i(k)\}}{\textcolor{red}{\sqrt{N}}} \] the measured static DNL/INL genuinely improves by \(\sqrt{N}\) at code-density (histogram) test

The composite result \[ \boxed{\sigma_{DNL,\max}^{TI} = \frac{\sigma_{\text{sub},DNL}}{\sqrt{N}}, \qquad \sigma_{INL,\max}^{TI} = \frac{\sigma_{\text{sub},INL}}{\sqrt{N}}} \] where \(N\) is channel number

Conventional binary array sub-SARADC \[ \sigma_{DNL,\max}=\sqrt{2^n-1}\,\frac{\sigma_u}{C_u}\,[\mathrm{LSB}], \qquad \sigma_{INL,\max}=\frac{\sqrt{2^n}}{2}\,\frac{\sigma_u}{C_u}\,[\mathrm{LSB}] \] VCM-based (\(n-1\) bit array) sub-SARADC

In a VCM-based (top-plate-sampled, tri-level) SAR, the MSB decision needs no capacitor switching at all — it's a direct comparison against \(V_{CM}\) \[ \sigma^{VCM}_{DNL,\max} = \sqrt{2^{n-1}-1}\,\frac{\sigma_u}{C_u}, \qquad \sigma^{VCM}_{INL,\max} = \frac{\sqrt{2^{n-1}}}{2}\,\frac{\sigma_u}{C_u} \qquad \text{at } k=2^{n-2},\,3\cdot 2^{n-2} \]


image-20260913173227688

reference

Poulton, Ken. ISSCC2009 "Time-Interleaved ADCs, Past and Future" (slides)

—. CICC2010 "GHz ADCs: From Exotic to Mainstream", tutorial session, (slides)

—. ISSCC2015 "Interleaved ADCs Through the Ages", (slides)

ISSCC2015 F1: High-Speed Interleaved ADCs

Ron Kapusta, Analog Devices, CICC2015 ED007: SAR ADCs in parallel [time-interleaved] converter arrays

Samuel Palermo, ISSCC 2018 T10: ADC-Based Serial Links: Design and Analysis

Ewout Martens. ESSCIRC 2019 Tutorials: Advanced Techniques for ADCs for 5G Massive MIMO [https://youtu.be/7hYichGGU6k]

Yohan Frans, CICC2019 ES3-3- "ADC-based Wireline Transceivers" [pdf]

John P. Keane, ISSCC2020 T5: "Fundamentals of Time-Interleaved ADCs"

Heng Zhang, Broadcom, ISSCC 2025 Forum F4.2: High-speed ADCs for 100Gbps+ Wireline Transceivers

Athanasios Ramkaj. January 26, 2022, IEEE SSCS Santa Clara Valley Section Technical Talk: Design Considerations Towards Optimal High-Resolution Wide-Bandwidth Time-Interleaved ADCs [video] [slides]

V. Chen, "Tutorial: High-Speed Analog-to-Digital Converters," 2025 IEEE International Solid-State Circuits Conference (ISSCC), San Francisco, CA, USA, 2025, pp. 1-1, doi: 10.1109/ISSCC49661.2025.11076112.


Ahmed M. A. Ali 2016, "High Speed Data Converters"

Razavi, B. (2025). Analysis and design of data converters. Cambridge University Press.

S. Jang, J. Lee, Y. Choi, D. Kim, and G. Kim, "Recent advances in ultra-high-speed wireline receivers with ADC-DSP-based equalizers," IEEE Open Journal of the Solid-State Circuits Society (OJ-SSCS), vol. 4, pp. 290-304, Nov. 2024.

Yida Duan. Design Techniques for Ultra-High-Speed Time-Interleaved Analog-to-Digital Converters (ADCs) [http://www2.eecs.berkeley.edu/Pubs/TechRpts/2017/EECS-2017-10.pdf]

Preview Lecture #1 - "Extreme SAR ADCs" Online Course (2024) - Prof. Chi-Hang Chan (U. of Macau) [https://youtu.be/rgMRL4QZ-wA]


oscarmattia. Data Converter Toolbox [https://github.com/oscarmattia/data_converter_toolbox]

image-20250703212349339


image-20250726092747147

VGA/attenuator: ensure a constant swing at the slicer input regardless of the channel variation


CTLE Linearity

TODO 📅

image-20260328182645140

Front-End Noise

https://people.engr.tamu.edu/spalermo/ecen689/lecture6_ee720_rx_circuits.pdf

image-20250726180326659

image-20250726181859362

Decision Feedback Equalizer (DFE)

speculative DFE is also known as loop unrolled DFE, which solve the critical timing on first tap

DFE architecture

image-20260328182441169

image-20250607235201147

Extensive work on DFEs has produced a multitude of architectures, which can be broadly categorized as "direct"" or "unrolled" (speculative) DFEs with "full-rate" or "half-rate" clocking

image-20250608000306928

image-20250608000338808

image-20250608000357010

S. Ibrahim and B. Razavi, "Low-Power CMOS Equalizer Design for 20-Gb/s Systems," in IEEE Journal of Solid-State Circuits, vol. 46, no. 6, pp. 1321-1336, June 2011 [https://sci-hub.se/10.1109/JSSC.2011.2134450]

S. Ibrahim and B. Razavi, Low-Power DFE Design [https://picture.iczhiku.com/resource/eetop/wykflwIuIQDzYNcB.PDF]

DFE Error Propagation

Geoff Zhang. Preliminary Studies on DFE Error Propagation, Precoding, and their Impact on KP4 FEC Performance for PAM4 Signaling Systems [https://www.ieee802.org/3/ck/public/18_09/zhang_3ck_01a_0918.pdf]

Cathy Liu, The Effect of DFE Error Propagation [https://www.ieee802.org/3/ap/public/nov05/liu_01_1105.pdf]

Yuchun Lu, Huawei, Elimination of DFE Error Propagation and Post-FEC Error Floor (Precoding 2.0) [https://www.ieee802.org/3/ck/public/19_03/lu_3ck_01_0319.pdf]

Z. Wu and J. -R. Guo, "Analysis of UCIe 48/64 GT/s Electrical Links," in IEEE Open Journal of the Solid-State Circuits Society [pdf]

TODO 📅

image-20260328182254717

PAM4 DFE

image-20250525202236767

image-20250525210606180

image-20250525221556845

image-20250525221432513

image-20250525221148218

K. -C. Chen, W. W. -T. Kuo and A. Emami, "A 60-Gb/s PAM4 Wireline Receiver With 2-Tap Direct Decision Feedback Equalization Employing Track-and-Regenerate Slicers in 28-nm CMOS," in IEEE Journal of Solid-State Circuits, vol. 56, no. 3, pp. 750-762, March 2021 [https://www.mics.caltech.edu/wp-content/uploads/2021/02/JSSC-2020-Xavier-PAM4-Receiver.pdf]

Hongtao Zhang, DesignCon 2016. PAM4 Signaling for 56G Serial Link Applications − A Tutorial [https://www.xilinx.com/publications/events/designcon/2016/slides-pam4signalingfor56gserial-zhang-designcon.pdf]

Edge DFE

K. -L. J. Wong, E. -H. Chen and C. -K. K. Yang, "Edge and Data Adaptive Equalization of Serial-Link Transceivers," in IEEE Journal of Solid-State Circuits, vol. 43, no. 9, pp. 2157-2169, Sept. 2008 [[https://sci-hub.ru/10.1109/JSSC.2008.2001876]*https://sci-hub.ru/10.1109/JSSC.2008.2001876]

B. Brunn, “Edge equalization NRZ,” Jul. 2004 [Online]. Available: [http://www.ieee802.org/3/ap/public/jul04/brunn_01_0704.pdf]

CC Chen, "Why Edge DFE?" [https://youtu.be/azkm7A9plyY]

TODO 📅

Peak to Main Ratio (PMR)

Yohan Frans, CICC2019 ES3-3- "ADC-based Wireline Transceivers" [pdf]

Boesch, et.al., “A 0.003 mm2 5.2 mW/tap 20 GBd inductor-less 5-tap analog RX-FFE,” 2016 IEEE Symposium on VLSI Circuits (VLSI-Circuits), 2016 [https://sci-hub.ru/10.1109/VLSIC.2016.7573522]

—, “Signal preconditioning using feedforward equalizers in ADC-based data links”, Ph.D. Dissertation, Stanford University, 2016 [https://purl.stanford.edu/dk653rc7126]

K. Zheng, “System-Driven Circuit Design for ADC-Based Wireline Data Links”, Ph.D. Dissertation, Stanford University, 2018 [https://purl.stanford.edu/hw458fp0168]

image-20261001074247470

image-20261024195325839

image-20261024195306253

image-20261024200924921

reference

T. Chan Carusone, T. O. Dickson, S. Palermo, S. Shekhar and M. Mansuri, "Modern Wireline Transceivers," in IEEE Journal of Solid-State Circuits, vol. 61, no. 2, pp. 395-422, Feb. 2026 [https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=11311714]

S. Jang, J. Lee, Y. Choi, D. Kim and G. Kim, "Recent Advances in Ultrahigh-Speed Wireline Receivers With ADC-DSP-Based Equalizers," in IEEE Open Journal of the Solid-State Circuits Society, vol. 4, pp. 290-304, 2024 [https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=10767763]

Miguel Gandara, MediaTek. CICC 2025 Circuit Insights: Basics of Wireline Receiver Circuits [https://youtu.be/X4JTuh2Gdzg]

Tony Chan Carusone, Alphawave Semi. VLSI2025 SC2: Connectivity Technologies to Accelerate AI

Noman Hai, Synopsys, Canada CASS Talks 2025 - May 2, 2025: High-speed Wireline Interconnects: Design Challenges and Innovations in 224G SerDes [https://www.youtube.com/live/wHNOlxHFTzY]

Nhat Nguyen and Masum Hossain, ISSCC 2021 Forum F6.7: 112Gb/s-and-Beyond Long-Reach and Short-Reach Electrical Interfaces

Jihwan Kim, Intel, ISSCC 2023 Forum F1.5: Circuit Designs for 200+Gb/s Electrical Transceivers

Ariel Cohen, Intel, ISSCC 2024 Forum F6.3: Beyond 200Gbps Electrical transceivers – Circuit Architecture, Design Implementation and Silicon Results

Heng Zhang, Broadcom, ISSCC 2025 Forum F4.2: High-speed ADCs for 100Gbps+ Wireline Transceivers

E-Hung Chen, MTK, ISSCC 2026 Forum F2.3: State-of-the-Art 200+ Gb/s Electrical and Optical Interconnects

image-20261024112822262

Analog-based TX vs. DSP/DAC TX

image-20261001085907790

SST vs. CML Driver

Z. Toprak-Deniz et al., "A 128-Gb/s 1.3-pJ/b PAM-4 Transmitter With Reconfigurable 3-Tap FFE in 14-nm CMOS," in IEEE Journal of Solid-State Circuits, vol. 55, no. 1, pp. 19-26, Jan. 2020 [https://sci-hub.st/10.1109/JSSC.2019.2939081]

Design Challenges Of High-Speed Wireline Transmitters [https://semiengineering.com/design-challenges-of-high-speed-wireline-transmitters/]

image-20261024001902143

image-20260531080813258

The source-series terminated (SST) drivers are more power-efficient than their current mode logic (CML) counterparts due to their lower termination power

differential output amplitude

\[ V_{ad,SST} = \frac{V_{DD}}{4R_T}\cdot 2R_T = \boxed{I_{DD}\cdot 2R_T} \qquad V_{ad,CML} = \frac{I_{DD}}{4}\cdot 2R_T = \boxed{\frac{1}{4} I_{DD}\cdot 2R_T} \]

To achieve the same differential output amplitude, CML topologies consume \(4\) times the current of SST topologies



image-20240825194548697

Current mode drivers become power competitive at very high data rates

  • Dynamic power consumption scales with frequency \(\Longrightarrow\) SST drivers lose power advantage

Serialization Approaches

Z. Toprak-Deniz et al., "A 128-Gb/s 1.3-pJ/b PAM-4 Transmitter With Reconfigurable 3-Tap FFE in 14-nm CMOS," in IEEE Journal of Solid-State Circuits, vol. 55, no. 1, pp. 19-26, Jan. 2020 [https://sci-hub.st/10.1109/JSSC.2019.2939081]

triple-stacked 4:1 n-type MUX

image-20260530063153678

tripstack4to1MUX.drawio

2-1 mux timing

mux2-1_timing.drawio

The circuit alone does not fix the bit order. It depends on which clock edge arrives first:

  • First edge rising: the a-flop captures first, and the output is a₀, b₀, a₁, b₁, …
  • First edge falling: the b-flop captures first, and the output is b₀, a₀, b₁, a₁, …

mux2-1_serializer_timing.drawio

The inverter chain sets the hold margin at the mux

  • The first two inverters delay the flop clocks behind the select, so each flop updates only after the mux has switched away from it

  • The third inverter mainly provides the inverted clock for the b-flop, and it adds a little extra hold margin to that path

Because t_su + t_hd ≈ T/2, every bit of delay added for hold comes out of setup. The chain should therefore be just long enough to guarantee positive hold margin across PVT.

divider latch timing

div2_mux_retime_latch.drawio

The clk_d inverters set where the latch's transparent window sits inside di's stable window:

  • d_div is the delay from the clk edge to di changing: t_inv + t_cq,FF + t_mux,sel.
  • d_c is the delay of the clk_d path: 2·t_inv, plus the latch's internal delay.

From those:

t_hd ≈ d_div − d_c t_su ≈ T/2 +d_c − d_div

The sum is fixed at T/2, so adding clk_d delay buys setup by spending hold, and removing it does the reverse. The window is centred (t_su = t_hd = T/4) when d_c ≈ d_div − T/4

Two latches

two_latch_half_period_shift_1.drawio

1-UI Data Stagger

C. Menolfi et al., "6.2 A 112Gb/S 2.6pJ/b 8-Tap FFE PAM-4 SST TX in 14nm CMOS," 2018 IEEE International Solid-State Circuits Conference - (ISSCC), San Francisco, CA, USA, 2018, pp. 104-106 [https://sci-hub.ru/10.1109/ISSCC.2018.8310205]

Z. Toprak-Deniz et al., "6.6 A 128Gb/s 1.3pJ/b PAM-4 Transmitter with Reconfigurable 3-Tap FFE in 14nm CMOS," 2019 IEEE International Solid-State Circuits Conference - (ISSCC), San Francisco, CA, USA, 2019, pp. 122-124 [https://sci-hub.ru/10.1109/ISSCC.2019.8662479]

—, "A 128-Gb/s 1.3-pJ/b PAM-4 Transmitter With Reconfigurable 3-Tap FFE in 14-nm CMOS," in IEEE Journal of Solid-State Circuits, vol. 55, no. 1, pp. 19-26, Jan. 2020 [https://sci-hub.ru/10.1109/JSSC.2019.2939081]

T. O. Dickson et al., "C3.2 A 72GS/s, 8-bit DAC-based Wireline Transmitter in 4nm FinFET CMOS for 200+Gb/s Serial Links," 2022 IEEE Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits), Honolulu, HI, USA, 2022, pp. 28-29 [https://sci-hub.ru/10.1109/VLSITechnologyandCir46769.2022.9830421]

—, "A 72-GS/s, 8-Bit DAC-Based Wireline Transmitter in 4-nm FinFET CMOS for 200+ Gb/s Serial Links," in IEEE Journal of Solid-State Circuits, vol. 58, no. 4, pp. 1074-1086, April 2023, doi: 10.1109/JSSC.2022.3228632

a.k.a Phase Aligner, Tap Delay Generator

image-20261001080205374


image-20261001151432789

quarter_rate_1ui_stagger.drawio

D4'<0> is launched by the C4 0° rising edge, <1> by 90°, <2> by 180° and <3> by 270°

Why the chains have 2, 3, 3 and 4 latches:

  • each latch passes data on to the next latch to open. That hop can only be 1 UI (a 90° step) or 2 UI (a 180° step)
  • the first 0° latch opens at 2 UI, and the outputs must launch at 4, 5, 6 and 7 UI.
    • one hop for <0> (+2)
    • two for <1> (+1 +2)
    • two for <2> (+2 +2)
    • three for <3> (+1 +2 +2)

These match the slide's chains, which are the shortest possible for those targets

The two hop types also have different margins, which ties back to the earlier figures:

  • 90° hop: the input changes in the middle of the next latch's hold phase, so setup and hold are both about 1 UI.
  • 180° hop: this is the master–slave case. The input changes right after the next latch closes, so hold margin is only clock-to-Q, while setup gets 2 UI

image-20261004170946701

a tap delay generator retime the incoming data and provide 1-UI-staggered quarter-rate data (D0-D3)

image-20261004184022619

tap_delay_gen_fig9-Fig. 9 as drawn (7 latches).drawio

Fig 9. works for D0–D2, but the D3 path races. So Fig. 9 is probably simplified

Fix: add one C4_Q latch to the D3 path, giving C4_I → C4_Q → C4_QB.

tap_delay_gen_fig9-D3 path fixed (8 latches).drawio

1-UI Pulse Generator

Noman Hai, Synopsys. CMOS Circuit Techniques for Wireline Transmitters [https://www.synopsys.com/webinars/wireline-transmitters-part-1.html]

1-UI Pulsegen Architectures:

  • Logic Gates Style
  • Transmission Gates Style
  • Triple, Double and Single Stack MUX
  • 2-Stages 1-UI Generation
  • Tail-less CML
  • Dynamic Logic Style
  • NOR-NAND

Logic Gates Style

F. Celik, A. Akkaya, A. Tajalli and Y. Leblebici, "A 32-Gb/s PAM-4 SST Transmitter With Four-Tap FFE Using High-Impedance Driver in 28-nm FDSOI," in IEEE Transactions on Very Large Scale Integration (VLSI) Systems, vol. 29, no. 6, pp. 1132-1140, June 2021, [https://sci-hub.ru/10.1109/TVLSI.2021.3068242]

P. -J. Peng, Y. -T. Chen, S. -T. Lai, C. -H. Chen, H. -E. Huang and T. Shih, "6.7 A 112Gb/s PAM-4 Voltage-Mode Transmitter with 4-Tap Two-Step FFE and Automatic Phase Alignment Techniques in 40nm CMOS," 2019 IEEE International Solid-State Circuits Conference - (ISSCC), San Francisco, CA, USA, 2019, pp. 124-126, [https://sci-hub.ru/10.1109/ISSCC.2019.8662361]

—, "A 112-Gb/s PAM-4 Voltage-Mode Transmitter With Four-Tap Two-Step FFE and Automatic Phase Alignment Techniques in 40-nm CMOS," in IEEE Journal of Solid-State Circuits, vol. 56, no. 7, pp. 2123-2131, July 2021 [https://sci-hub.ru/10.1109/JSSC.2020.3038818]

image-20261108211330206

image-20261108211230731

image-20261108221019436

image-20261108221343619

Transmission Gates Style

C. Menolfi et al., "6.2 A 112Gb/S 2.6pJ/b 8-Tap FFE PAM-4 SST TX in 14nm CMOS," 2018 IEEE International Solid-State Circuits Conference - (ISSCC), San Francisco, CA, USA, 2018, pp. 104-106 [https://sci-hub.ru/10.1109/ISSCC.2018.8310205]

T. O. Dickson et al., "C3.2 A 72GS/s, 8-bit DAC-based Wireline Transmitter in 4nm FinFET CMOS for 200+Gb/s Serial Links," 2022 IEEE Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits), Honolulu, HI, USA, 2022, pp. 28-29 [https://sci-hub.ru/10.1109/VLSITechnologyandCir46769.2022.9830421]

—, "A 72-GS/s, 8-Bit DAC-Based Wireline Transmitter in 4-nm FinFET CMOS for 200+ Gb/s Serial Links," in IEEE Journal of Solid-State Circuits, vol. 58, no. 4, pp. 1074-1086, April 2023, doi: 10.1109/JSSC.2022.3228632

image-20261109002159240

image-20261108225611269

Dual passgate logic (DPL) \(\text{OR}\) and \(\text{AND}\) gates derive one-UI clock pulses from the quadrature C4 clocks

\(\text{C4x}=1\) and \(\text{C4X} + 90^\circ=0\) select the current transmission gate

image-20261004153931893


J. Kim et al., “A 224Gb/s DAC-Based PAM-4 Transmitter with 8-Tap FFE in 10nm CMOS,” ISSCC 2021 [https://sci-hub.jp/10.1109/ISSCC42613.2021.9365840]

—, "A 224-Gb/s DAC-Based PAM-4 Quarter-Rate Transmitter With 8-Tap FFE in 10-nm FinFET," in IEEE Journal of Solid-State Circuits, vol. 57, no. 1, pp. 6-20, Jan. 2022 [https://sci-hub.ru/10.1109/JSSC.2021.3108969]

duty correction & delay adjustment

image-20260921231822433

image-20260921231842187

Full-rate TX architecture

Sam Palermo, ECEN720: High-Speed Links Circuits and Systems Spring 2025 Lecture 5: Termination, TX Driver, & Multiplexer Circuits [https://people.engr.tamu.edu/spalermo/ecen689/lecture5_ee720_termination_txdriver.pdf]

J. Cao et al., "OC-192 transmitter and receiver in standard 0.18-/spl mu/m CMOS," in IEEE Journal of Solid-State Circuits, vol. 37, no. 12, pp. 1768-1780, Dec. 2002 [https://sci-hub.jp/10.1109/JSSC.2002.804336]

image-20261024143217254

With the FFs, latches, and clocks unchanged, reversing the MUX selection still works, but adds latency

The bit order is preserved; each bit is selected later.

  • Reversing both first-stage MUXes adds 2 UI
  • Reversing the final MUX adds 1 UI.
  • Reversing all three preserves (D_0,D_1,D_2,D_3,), with 3 UI additional latency

image-20261001145947746

The retimer between the final stage of the MUX and the output driver is used to reduce the data jitter due to the bandwidth limitation of the selection circuit in the 2 : 1 MUX cell and duty cycle distortion of the half-rate clock driving that stage


image-20261024162436538

Half-rate TX architecture

M. Meghelli et al., "A 10Gb/s 5-Tap-DFE/4-Tap-FFE Transceiver in 90nm CMOS," 2006 IEEE International Solid State Circuits Conference - Digest of Technical Papers, San Francisco, CA, USA, 2006, pp. 213-222 [https://sci-hub.ru/10.1109/ISSCC.2006.1696051]

J. F. Bulzacchelli et al., "A 10-Gb/s 5-Tap DFE/4-Tap FFE Transceiver in 90-nm CMOS Technology," in IEEE Journal of Solid-State Circuits, vol. 41, no. 12, pp. 2885-2900, Dec. 2006 [https://sci-hub.ru/10.1109/JSSC.2006.884342]

Yang, Chih-Kong Ken. Design of high-speed serial links in CMOS. Stanford University, 1999. [http://i.stanford.edu/pub/cstr/reports/csl/tr/98/775/CSL-TR-98-775.pdf]

Mark Horowitz, Chih-Kong Ken Yang, and Stefanos Sidiropoulos. 1998. High-Speed Electrical Signaling: Overview and Limitations. IEEE Micro 18, 1 (January 1998), 12–24. https://doi.org/10.1109/40.653013 [https://people.engr.tamu.edu/spalermo/ecen689/hs_electrical_signaling_horowitz_micro_1998.pdf]

Five-latch 2:1 MUX

image-20261024004405229

image-20261021232016291

halfrate_tamu.drawio


image-20261024162210709

image-20261024162248410

halfrate_swap.drawio


image-20261001111219814

image-20261001170626200

The half period that second-half selection "wastes" is deliberate slack: it lets each input settle fully before it is passed. You're trading a little latency for robustness, and designers almost always take that trade. If latency truly mattered, the better move would be to trim pipeline stages or the FIFO depth elsewhere, not to remove the settling slack from the highest-speed MUX.

Quarter-rate TX architecture

Z. Toprak-Deniz et al., "6.6 A 128Gb/s 1.3pJ/b PAM-4 Transmitter with Reconfigurable 3-Tap FFE in 14nm CMOS," 2019 IEEE International Solid-State Circuits Conference - (ISSCC), San Francisco, CA, USA, 2019, pp. 122-124 [https://sci-hub.ru/10.1109/ISSCC.2019.8662479]

—, "A 128-Gb/s 1.3-pJ/b PAM-4 Transmitter With Reconfigurable 3-Tap FFE in 14-nm CMOS," in IEEE Journal of Solid-State Circuits, vol. 55, no. 1, pp. 19-26, Jan. 2020 [https://sci-hub.ru/10.1109/JSSC.2019.2939081]

Quarter-Rate: A clocking or sampling architecture where the internal circuit clock runs at one-fourth (1/4) of the total serial data rate

Quadrature: A relationship between two signals or clocks that have a 90o phase difference (a quarter of a complete wave cycle), commonly used for I/Q modulation, directional tracking in encoders, or generating multi-phase clocks

quadrature quarter-rate (C4)

image-20261001154407300

image-20261001154438645


Fig. 5(c): The 2-UI pulse D1′ is carved by C4IB alone — it starts on C4IB rising and ends on C4IB falling. For the D1 → D1′ stage, the margins are 1.5 UI before and 0.5 UI after, which is asymmetric

Fig. 5(d): D1′ is the 1-UI pulse, C4IB isn't the only reference — It starts on C4IB rising, but it ends on C4Q falling, as the arrows in the figure show. The pulse generator is enabled only while C4IB and C4Q are both high

image-20261001200238448

Case Window D1 must be stable over Before After
(c) D1 → D1′ C4IB high (2 UI) 1.5 UI 0.5 UI
(c) D1 → D_OP C4IB high and C4Q high (1 UI) 1.5 UI 1.5 UI
(d) D1 → D1′ C4IB high and C4Q high (1 UI) 1 UI 2 UI

The 0.5 UI in Fig. 5(c) is an idealized drawing, not a real delay value. In silicon, the D1 edge occurs at the launching C4 edge plus the latch clock-to-Q delay plus wiring delay. The authors drew it at 0.5 UI to show the ideal centered placement with symmetric margin

The two sub-figures place D1 differently, which shows the data-to-clock offset is set by design and illustration choices. The real requirement is only that D1 is stable, with margin, whenever its carving gate is enabled

If the natural delay lands too close to an active edge, the designer can fix it by choosing a different launching clock phase or adding delay


image-20261108092947736

image-20261108092820041

image-20261108093731862


A. A. Hafez, M. -S. Chen and C. -K. K. Yang, "A 32-to-48Gb/s serializing transmitter using multiphase sampling in 65nm CMOS," 2013 IEEE International Solid-State Circuits Conference Digest of Technical Papers, San Francisco, CA, USA, 2013, pp. 38-39 [https://sci-hub.ru/10.1109/ISSCC.2013.6487627]

P. -J. Peng, Y. -T. Chen, S. -T. Lai, C. -H. Chen, H. -E. Huang and T. Shih, "6.7 A 112Gb/s PAM-4 Voltage-Mode Transmitter with 4-Tap Two-Step FFE and Automatic Phase Alignment Techniques in 40nm CMOS," 2019 IEEE International Solid-State Circuits Conference - (ISSCC), San Francisco, CA, USA, 2019, pp. 124-126, [https://sci-hub.ru/10.1109/ISSCC.2019.8662361]

image-20261108214551856

The timing examples below use 1 UI = one output symbol: the 14-GHz clocks have a 4-UI period, and the 7-GHz clocks have an 8-UI period. For illustration, CK0,14G rises at 0 UI and CK0,7G rises at −1 UI; the 8:4 mux and each latch have a 0.25-UI propagation delay. The 1-UI select pulses are the overlaps of adjacent 14-GHz clock phases, occupying the latter half of each high phase. Dout shows ideal 4:1 selection, omitting the final mux's propagation delay.

8 latches — simultaneous 8:4 outputs, followed by four master–slave pairs

Quarter-rate 4-to-1 mux timing with eight latches

M0–M3 are transparent while their clocks are low; Q0–Q3 are transparent while their clocks are high. The master latches prevent a new simultaneous input word from propagating through the later lanes while those lanes are selected.

4 latches — staggered 8:4 outputs from an octal-phase divider

Quarter-rate 4-to-1 mux timing with four latches and an octal-phase divider

The 7-GHz phases 0°, 45°, 90°, 135° stagger R0–R3 by 0, 1, 2, 3 UI; their complementary phases complete the octal clock set. Each lane now needs only one high-transparent latch. Both examples produce the same a, b, c, d, … output, with the selected data stable throughout every 1-UI pulse.

Synchronized divider

M. A. Kossel et al., "8.3 An 8b DAC-Based SST TX Using Metal Gate Resistors with 1.4pJ/b Efficiency at 112Gb/s PAM-4 and 8-Tap FFE in 7nm CMOS," 2021 IEEE International Solid-State Circuits Conference (ISSCC), San Francisco, CA, USA, 2021, pp. 130-132 [https://sci-hub.ru/10.1109/ISSCC42613.2021.9365784]

—. (2023), Design of synchronous frequency dividers in 5-nm FinFET CMOS technology. Electron. Lett., 59: e13033. [https://ietresearch.onlinelibrary.wiley.com/doi/epdf/10.1049/ell2.13033]

Michael Perrott August 12, 2008, Short Course On Phase-Locked Loops and Their Applications Day 2, PM Lecture Basic Building Blocks (Part II) High Speed Frequency Dividers, Phase Detectors, Charge Pumps, and Loop Filter Design [https://cppsim.org/PLL_Lectures/day2_pm.pdf]

Q. J. Gu and Z. Gao, "A CMOS High Speed Multi-Modulus Divider With Retiming for Jitter Suppression," in IEEE Microwave and Wireless Components Letters, vol. 23, no. 10, pp. 554-556, Oct. 2013, doi: 10.1109/LMWC.2013.2248080 [https://www.ece.ucdavis.edu/hsics/wp-content/uploads/sites/4/2015/04/2013_A-CMOS-High-Speed-Multi-Modulus-Divider-With-Retiming-for-Jitter-Suppression.pdf]

image-20261001082434120

The lower speed sub-rate clocks are then obtained using a synchronous divider based on conventional master-slave flip-flops

syndiv8

syndiv8_wv.drawio

The preceding synchronous divider is equivalent to the synchronous implementation described below

image-20261001085428000

Each stage's toggle decision is computed from the states of all previous stages, but its timing comes only from the common input clock

SST Driver

sharing termination in SST transmitter

tx_leg.drawio

Sharing termination keep a constant current through leg, which improve TX speed in this way. On the other hand, the sharing termination facilitate drain/source sharing technique in layout.

pull-up and pull-down resistor

sst-evolution

Original stacked structure

Pro's:

​ smaller static current when both pull up and pull down path is on

Con's:

​ slowly switching due to parasitic capacitance behind pull-up and pull-down resistor

with single shared linearization resistor

Pro's:

​ The parasitic capacitance behind the resistor still exists but is now always driven high or low actively

Con's:

​ more static current

VM Driver Equalization - differential ended termination

\[ V_o = D_{n+1}C_{-1}+D_nC_0+D_{n-1}C_{+1} \]

where \(D_n \in \{-1, 1\}\)

vdrv.drawio \[ V_{\text{rx}} = V_{\text{dd}} \frac{(R_2-R_1)R_T}{R_1R_T+R_2R_T+R_1R_2} \] With \(\color{blue}R_u=(L+M+N)R_T\)

Normalize above equation, obtain \[ V_{\text{rx,norm}} = \frac{(R_2-R_1)R_T}{R_1R_T+R_2R_T+R_1R_2} \]

\(D_{n-1}\) \(D_{n}\) \(D_{n+1}\)
\(C_{-1}\) 1 -1 -1
\(C_0\) -1 1 -1
\(C_{+1}\) -1 -1 1

Where precursor \(R_L = L\times R_T\), main cursor \(R_M = M\times R_T\) and post cursor \(R_N = N\times R_T\)

image-20220709151054840

Equation-1

\(D_{n-1}D_nD_{n+1}=1,-1,-1\)

pre.drawio

\[ R_1 = R_N = \frac{R_u}{N} \qquad\qquad R_2 = R_L\parallel R_M = \frac{R_u}{L+M} \]

We obtain \[ V_{L}= \frac{1}{2}\cdot\frac{N-(L+M)}{L+M+N} \]

Equation-2

\(D_{n-1}D_nD_{n+1}=-1,1,-1\)

main.drawio

with \(R_1=R_T\) and \(R_2=+\infty\), we obtain \[ V_M = \frac{1}{2} \]

Equation-3

\(D_{n-1}D_nD_{n+1}=-1,-1,1\)

\[ R_1 = R_L = \frac{R_u}{L} \qquad R_2 = R_N\parallel R_M = \frac{R_u}{N+M} \]

We obtain \[ V_N = \frac{1}{2}\cdot\frac{L-(N+M)}{L+M+N} \]

Obtain FIR coefficients

We define \[ l = \frac{L}{L+M+N} \qquad m = \frac{M}{L+M+N} \qquad n = \frac{N}{L+M+N} \]

where \(l+m+n=1\)

Due to Eq1 ~ Eq3 \[ \left\{ \begin{array}{cl} C_{-1}-C_0-C_1 & = \frac{1}{2}(n-l-m) \\ -C_{-1}+C_0-C_1 & = \frac{1}{2} \\ -C_{-1}-C_0+C_1 & = \frac{1}{2}(l-n-m) \end{array} \right. \] After scaling, we get \[ \left\{ \begin{array}{cl} C_{-1}-C_0-C_1 & = -l-m+n \\ -C_{-1}+C_0-C_1 & = l+m+n \\ -C_{-1}-C_0+C_1 & = l-m-n \end{array} \right. \] Then, the relationship between FIR coefficients and legs is clear, i.e. \[ C_{-1} = -\frac{L}{L+M+N} \qquad C_{0} = \frac{M}{L+M+N} \qquad C_{1} = -\frac{N}{L+M+N} \]

For example, \(C_{-1}=-0.1\), \(C_0=0.7\) and \(C_1=-0.2\) \[ H(z) = -0.1+0.7z^{-1}-0.2z^{-2} \] image-20220709185832444

1
2
3
4
5
6
7
w = [-0.1, 0.7, -0.2];
Fs = 32e9;
[mag, w] = freqz(w, 1, [], Fs);
plot(w/1e9, abs(mag));
xlabel('Freq(GHz)');
ylabel('mag');
grid on;

VM Driver Equalization - single ended termination

Equation-1

pre_se.drawio

\[ V_{\text{rxp}} = \frac{1}{2} \cdot \frac{N}{L+M+N} \qquad V_{\text{rxm}} = \frac{1}{2} \cdot \frac{L+M}{L+M+N} \] So \[ V_{L}= \frac{1}{2}\cdot\frac{N-(L+M)}{L+M+N} \] which is same with differential ended termination

Equation-2

main_se.drawio

\[ V_{\text{rxp}} = \frac{1}{2} \qquad V_{\text{rxm}} = 0 \]

So \[ V_{M}= \frac{1}{2} \] which is same with differential ended termination

Equation-3 \[ V_{N}= \frac{1}{2}\cdot\frac{L-(N+M)}{L+M+N} \]

Obtain FIR coefficients

Same with differential ended termination driver.

Tailless CML driver

G. Steffan et al., "6.4 A 64Gb/s PAM-4 transmitter with 4-Tap FFE and 2.26pJ/b energy efficiency in 28nm CMOS FDSOI," 2017 IEEE International Solid-State Circuits Conference (ISSCC), San Francisco, CA, USA, 2017, pp. 116-117 [https://sci-hub.ru/10.1109/ISSCC.2017.7870288]

Z. Toprak-Deniz et al., "A 128-Gb/s 1.3-pJ/b PAM-4 Transmitter With Reconfigurable 3-Tap FFE in 14-nm CMOS," in IEEE Journal of Solid-State Circuits, vol. 55, no. 1, pp. 19-26, Jan. 2020 [https://sci-hub.ru/10.1109/JSSC.2019.2939081]

J. Im et al., "6.1 A 112Gb/s PAM-4 Long-Reach Wireline Transceiver Using a 36-Way Time-Interleaved SAR-ADC and Inverter-Based RX Analog Front-End in 7nm FinFET," 2020 IEEE International Solid-State Circuits Conference - (ISSCC), San Francisco, CA, USA, 2020, pp. 116-118 [https://sci-hub.ru/10.1109/ISSCC19947.2020.9063081]

—, "A 112-Gb/s PAM-4 Long-Reach Wireline Transceiver Using a 36-Way Time-Interleaved SAR ADC and Inverter-Based RX Analog Front-End in 7-nm FinFET," in IEEE Journal of Solid-State Circuits, vol. 56, no. 1, pp. 7-18, Jan. 2021 [https://sci-hub.ru/10.1109/JSSC.2020.3024261]

a.k.a. Pseudo - CML

image-20261021234731893

image-20261001180040383

image-20261017095055911

image-20261020001618501

Active Peaking CMOS Pre-Driver

C. Menolfi et al., "A 112Gb/S 2.6pJ/b 8-Tap FFE PAM-4 SST TX in 14nm CMOS," 2018 IEEE International Solid-State Circuits Conference - (ISSCC), San Francisco, CA, USA, 2018, pp. 104-106 [https://sci-hub.ru/10.1109/ISSCC.2018.8310205]

HungWen Lu, ChauChin Su and Chien-Nan Liu, "A scalable digitalized buffer for gigabit I/O," 2008 IEEE Custom Integrated Circuits Conference, San Jose, CA, USA, 2008, pp. 241-244 [https://sci-hub.ru/10.1109/CICC.2008.4672068]

image-20261001072415008

active-peaking \[ I_\text{x} = g_\text{mfb}\cdot V_\text{x}\frac{1}{sR_\text{tg}C_\text{g}+1} \qquad \Longrightarrow \qquad R_\text{x}=\frac{V_\text{x}}{I_\text{x}}= \frac{1}{g_\text{mfb}} + s\frac{R_\text{tg}C_\text{g}}{g_\text{mfb}} \]

i.e. \[ R \approx \frac{1}{g_\text{mfb}} \qquad\qquad L \approx \frac{R_\text{tg}C_\text{g}}{g_\text{mfb}} \]

image-20251217231902887


U. Singh et al., "8.6 A 280mW 112Gb/s PAM-4/NRZ Transceiver for Low-Power IOs in 5nm FinFET Technology," 2026 IEEE International Solid-State Circuits Conference (ISSCC), San Francisco, CA, USA, 2026, pp. 144-146, doi: 10.1109/ISSCC49663.2026.11409174

K. Thasari et al., "A 280-mW, 112-Gb/s PAM-4/NRZ Transceiver for Low-Power IOs in 5-nm FinFET Technology," in IEEE Journal of Solid-State Circuits, doi: 10.1109/JSSC.2026.3743919

image-20261026213214551

Single-Ended-to-Differential (S2D)

T. Dickson et al., "C3.2 A 72GS/s, 8-bit DAC-based Wireline Transmitter in 4nm FinFET CMOS for 200+Gb/s Serial Links," 2022 IEEE Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits), Honolulu, HI, USA, 2022, pp. 28-29 [https://sci-hub.ru/10.1109/VLSITechnologyandCir46769.2022.9830421]

image-20261001151302086

FFE w/ Quarter-rate Clocks in Analog-TX

J. Kim et al., "6.1 A 112Gb/s PAM-4 transmitter with 3-Tap FFE in 10nm CMOS," 2018 IEEE International Solid-State Circuits Conference - (ISSCC), San Francisco, CA, USA, 2018, pp. 102-104, [https://sci-hub.ru/10.1109/ISSCC.2018.8310204]

— "A 112 Gb/s PAM-4 56 Gb/s NRZ Reconfigurable Transmitter With Three-Tap FFE in 10-nm FinFET," in IEEE Journal of Solid-State Circuits, vol. 54, no. 1, pp. 29-42, Jan. 2019 [https://sci-hub.ru/10.1109/JSSC.2018.2874040]

image-20261108104105794

Fig. 5 has inconsistent symbol labels with Fig. 4(c)

image-20261108110840310

trade latency with extra setup margin

Quarter-rate FFE timing with three 3-UI retiming steps

The highlighted path is D<1> → D<1>' → D<1>" → post<1>, with 3 UI between successive sampling edges

Basic serializer unit — Fig. 4(b)

image-20261108184523246

Basic serializer timing: divided clock, internal registers, mux, and output capture

Q0 is the upper flip-flop output; Q1a and Q1b are the first and second lower flip-flop outputs. Q0 and Q1a capture on ↑CKdiv, while Q1b captures on ↓CKdiv. The mux selects Q0 when CKdiv = 0 and Q1b when CKdiv = 1.

Delays are illustrative. The output flip-flop samples on ↑CK, before the delayed divider transition changes the mux selection. The available setup interval is therefore \(T_\mathrm{CK}-t_\mathrm{div}-t_\mathrm{mux}\):

\[ t_\mathrm{div}+t_\mathrm{mux}+t_\mathrm{su}<T_\mathrm{CK} \]

3-UI timing path at interface

image-20261108180223263

Pulse Generator RJ Filtering

J. Kim et al., “8.1 A 224Gb/s DAC-Based PAM-4 Transmitter with 8-Tap FFE in 10nm CMOS,” ISSCC 2021 [https://sci-hub.jp/10.1109/ISSCC42613.2021.9365840]

—, "A 224-Gb/s DAC-Based PAM-4 Quarter-Rate Transmitter With 8-Tap FFE in 10-nm FinFET," in IEEE Journal of Solid-State Circuits, vol. 57, no. 1, pp. 6-20, Jan. 2022 [https://sci-hub.ru/10.1109/JSSC.2021.3108969]

TODO 📅

DSP/Seralizer boundary

Synopsys PHY IP PCIe7 PHY x4 for TSMC 3PFF Databook, dwc_pcie7phy_tsmc3pff_x4ns — Product Code: J132-0, PHY Version: 1.05a February 23, 2026 [https://bbs.eetop.cn/thread-996408-1-1.html]

image-20261024010318929

image-20261117234856759

image-20261117235124021

word clock

TODO 📅

Divider reset synchronizer

TODO 📅

Lane-to-Lane Output Skew & clock shift

T. Pontius, "Alignment and deskew for multiple lanes of serial interconnect," WO Patent WO2007052229A2, May 10, 2007. [Online]. [https://patents.google.com/patent/WO2007052229A2/en]

P. Novellini and G. Guasti, "Lane-to-lane de-skew for transmitters," U.S. Patent 9 331 724 B2, May 3, 2016. [Online]. [https://patents.google.com/patent/US9331724B2/en]

70869 - TX Lane to Lane Alignment [https://adaptivesupport.amd.com/s/article/70869?language=en_US]

After reset, each channel latency is unknown. In a multi-lane link, the residual skew after reset can be many UIs.

The TX BUFFER, that was originally removed for minimal and deterministic latency, is the key to guaranteeing the known latency and thus the lane deskew

The writing and reading pointer positions are constantly measured, and when some particular conditions happen (buffer half full, buffer overflow, buffer empty ..) a flag raises.

image-20261117220548864

image-20261117221603878

suppose READING CLOCK is word clock

By driving the TX Phase Interpolator, we can delay or increment the phase of the reading pointer. This is represented by Circuit 1 in the picture below.

Thus, by controlling the buffer state of fullness it is possible to precisely control the latency in the single channel and as a consequence, the skew between lanes.

image-20261117222016396


PCI Express® Base Specification Revision 7.0 June 05, 2025 Table 8-6 Data Rate Dependent Transmitter Parameters

image-20261117230437385

Peak Power Constraint of TX FIR

Kevin Zheng , Circuit Insights @ ISSCC2025: Circuits for Wireline Communications [https://youtu.be/8NZl81Dj45M&t=829]

image-20250514215647905

Due to circuit limitation, circuit cannot have arbitrarily large voltage on the output, i.e. a limited maximum swing. In order to create the high frequency shape, the best we can do is lower DC gain (low frequency gain < 1)

  • FIR is not increasing the amplitude on the edges
  • FIR is reducing the inner eye diagram

The maximum swing stays the same, \(\sum_i |c_i|=1\)


0.5b in segmented DAC

Z. Guo et al., "6.2 A 112.5Gb/s ADC-DSP-Based PAM-4 Long-Reach Transceiver with >50dB Channel Loss in 5nm FinFET," 2022 IEEE International Solid-State Circuits Conference (ISSCC), San Francisco, CA, USA, 2022, pp. 116-118 [https://sci-net.xyz/10.1109/ISSCC42614.2022.9731650]

image-20261024205151890

image-20261024203740807

totalX = 1 +2 + 4 + 8 + (1 + 2 + 4)*16 + 4*16 = 1 +2 + 4 + 8 + 16*11 = 191

Therefore, the valid input range range is 0 ~ 191, even though the input bit number is 8 (range 0~255)

If we apply \(\log_2\) to \(\boxed{N+0.5}\) bit DAC \[ \log_2 (2^N + 2^{N-1}) = N - 1 + 3\log_2(3) = N + 0.585 \]

Receiver Detection

Synopsys PHY IP PCIe7 PHY x4 for TSMC 3PFF Databook, dwc_pcie7phy_tsmc3pff_x4ns — Product Code: J132-0, PHY Version: 1.05a February 23, 2026 [https://bbs.eetop.cn/thread-996408-1-1.html]

image-20261118000329319

FeedForward Equalization Theory

image-20220709111229772

image-20220709112543338

image-20220709125046329

Pre-cursor FFE can compensate phase distortion through the channel

image-20220709130050057

Single-ended termination

Differential termination

PAM4 TX

image-20220717010007963

Here, \(d_{\text{LSB}} \in \{-1, 1\}\), \(d_{\text{MSB}} \in \{-2, 2\}\) and \(d' \in \{ -3, -1, 1, 3 \}\)

Implementation-1 could potentially experience performance degradation due to

  1. Clock skew, \(\Delta t\), could make the eye misaligned horizontally
  2. Gain mismatch, \(\Delta G\), could cause eye nonlinearity
  3. Bandwidth mismatch, \(\Delta f_{\text{BW}}\), could make the eye misaligned vertically

image-20220717011129124

Typically, a 3-tap FIR (pre + main + post) TX de-emphasis is used

3-tap FIR results in \(4^3 = 64\) possible distinct signal levels

msb_lsb.drawio

\[\begin{align} R_U^M \parallel R_D^M &= \frac{3R_T}{2}\\ R_U^L \parallel R_D^L &= 3R_T \end{align}\]

Thevenin Equivalent Circuit is thevenin_1.drawio

Which can be simpified as thevenin_2.drawio \[\begin{align} V_{\text{rx}} &= \frac{1}{2}(V_p - V_m) \\ &= \frac{1}{2}(\frac{2}{3}(2V_{\text{MSB}}+V_{\text{LSB}})-1) \\ &=\frac{1}{3}(2V_{\text{MSB}}+V_{\text{LSB}})-\frac{1}{2} \end{align}\]

The above eqations demonstrate that the output \(V_{\text{rx}}\) is the linear sum of MSB and LSB; LSB and MSB have relative weight, i.e. 1 for LSB and 2 for MSB.

Assume pre cusor has \(L\) legs, main cursor \(M\) legs and post cursor \(N\) legs, which is same with the convention in "Voltage-Mode Driver Equalization"

The number of legs connected with supply can expressed as \[ n_{up} = (1-d_{n+1})L + d_{n}M + (1-d_{n-1})N \] Where \(d_n \in \{0, 1\}\), or \[ n_{up} = \frac{1}{2}(-D_{n+1}+1)L + \frac{1}{2}(D_{n}+1)M + \frac{1}{2}(-D_{n-1}+1)N \] Where \(D_n \in \{-1, +1\}\)

Then the number of legs connected with ground is \[ n_{dn}=L+M+N-n_{up} \] where \(n_{up}+n_{dn}=L+M+N\)

Voltage resistor divider \[\begin{align} V_o &= \frac{\frac{R_{U}}{n_{dn}}}{\frac{R_U}{n_{dn}}+\frac{R_U}{n_{up}}} \\ &= \frac{1}{2}- \frac{1}{2}D_{n+1}\frac{L}{L+M+N}+ \frac{1}{2}D_{n}\frac{M}{L+M+N}-\frac{1}{2}D_{n-1}\frac{N}{L+M+N} \\ &= \frac{1}{2}-\frac{1}{2}D_{n+1}\cdot l+ \frac{1}{2}D_{n}\cdot m-\frac{1}{2}D_{n-1}\cdot n \end{align}\]

where \(l+m+n=1\)

\(V_{\text{MSB}}\) and \(V_{\text{LSB}}\) can be obtained

\[\begin{align} V_{\text{MSB}} &= \frac{1}{2}-\frac{1}{2}D^{\text{MSB}}_{n+1}\cdot l+ \frac{1}{2}D^{\text{MSB}}_{n}\cdot m-\frac{1}{2}D^{\text{MSB}}_{n-1}\cdot n \\ V_{\text{LSB}} &= \frac{1}{2}-\frac{1}{2}D^{\text{LSB}}_{n+1}\cdot l+ \frac{1}{2}D^{\text{LSB}}_{n}\cdot m-\frac{1}{2}D^{\text{LSB}}_{n-1}\cdot n \end{align}\]

Substitute the above equation into \(V_{\text{rx}}\), we obtain the relationship between driver legs and FFE coefficients

\[\begin{align} V_{\text{rx}} &=\frac{1}{3}(2V_{\text{MSB}}+V_{\text{LSB}})-\frac{1}{2} \\ &= \frac{1}{3} \left\{ 2\left( \frac{1}{2}-\frac{1}{2}D^{\text{MSB}}_{n+1}\cdot l+ \frac{1}{2}D^{\text{MSB}}_{n}\cdot m- \frac{1}{2}D^{\text{MSB}}_{n-1}\cdot n \right) + \left( \frac{1}{2}-\frac{1}{2}D^{\text{LSB}}_{n+1}\cdot l+ \frac{1}{2}D^{\text{LSB}}_{n}\cdot m- \frac{1}{2}D^{\text{LSB}}_{n-1}\cdot n \right) \right\}-\frac{1}{2} \\ &= \left(-\frac{l}{6} \cdot 2 \cdot D^{\text{MSB}}_{n+1}+ \frac{m}{6} \cdot 2 \cdot D^{\text{MSB}}_{n}- \frac{n}{6} \cdot 2 \cdot D^{\text{MSB}}_{n-1}\right) + \left(-\frac{l}{6} \cdot D^{\text{LSB}}_{n+1}+ \frac{m}{6} \cdot D^{\text{LSB}}_{n}- \frac{n}{6} \cdot D^{\text{LSB}}_{n-1}\right) \\ &= -\frac{l}{6}(2 \cdot D^{\text{MSB}}_{n+1}+D^{\text{LSB}}_{n+1})+ \frac{m}{6}(2\cdot D^{\text{MSB}}_{n}+D^{\text{LSB}}_{n}) -\frac{n}{6}(2\cdot D^{\text{MSB}}_{n-1}+D^{\text{LSB}}_{n-1}) \end{align}\]

After scaling, we obtain \[ V_{\text{rx}} = -l\cdot(2 \cdot D^{\text{MSB}}_{n+1}+D^{\text{LSB}}_{n+1})+ m\cdot(2\cdot D^{\text{MSB}}_{n}+D^{\text{LSB}}_{n}) - n \cdot(2\cdot D^{\text{MSB}}_{n-1}+D^{\text{LSB}}_{n-1}) \] Where \(C_{-1} = l\), \(C_0 = m\) and \(C_{1}=n\), which is same with that of NRZ

Eye Linearity vs. RLM (Relative Level Mismatch)

Chaowaroj (Max) Wanotayaroj. Introduction to PAM4 [https://indico.cern.ch/event/979659/contributions/4127016/attachments/2159338/3642883/PAM4Eval%20-%20Dec2020%20Seminar.pdf]

TODO 📅

Tx Measurements

PAM4 Transmitter Test Challenges [https://harrisburg.psu.edu/files/pdf/16861/2019/05/06/tektronix_penn_state_si_april_12_2019.pdf]

PAM4 Signaling in High Speed Serial Technology: Test, Analysis, and Debug [https://download.tek.com/document/55W_60273_1_HR_Letter.pdf]

PCIe 7.0 Introduction PCIe 6.0 Anritsu/Tektronix Solution [https://map-assets.tek.com/map-assets/emea/pdf-files/PCIe7_0_Intro_PCIe_6_0_Solution.pdf]

Mike Hertz, Teledyne LeCroy: WEBINAR PAM4 Analysis and Measurement Considerations

Brandon Gore, Samtec, DesignCon 2025, Transmitter Power Spectral Density Noise Impact for 200 Gb/s PAM 4 per lane [pdf] [slides]

TODO 📅

TX Jitter Measurement

PCI-SIG, Update-on-PCIe8p0-Scope-bandwidth-study-and-jitter-measurement-Intel-2025-12-18_v3

image-20260426211026901

Linear Fit Pulse Response (LFPR)

Hsinho Wu, Intel. DesignCon 2021: SNDR Analysis & Its Impacts on Link Performance

Christiaan Bil (Intel), DesignCon 2026. An Experimental Study of PCIe Transmitter Equalization Preset Measurement Methods for 64 and 128 GT/s PAM4 Signaling

Dhruv Gupta, DesignCon 2026. PAM4 measurements through lossy channels – why oscilloscope CDR emulation matters

TODO 📅

SNDR

Marianne Nourzad, July 2nd, 2020 PCI-SIG ® EWG Meeting, PCIE Gen6 TX SNDR Methodology Discussion

Pegah Alavi (Keysight Technologies) DesignCon 2025: PCI Express & PAM4: Balancing Silicon and interconnect interdependencies for 128 GT/s

Rick Eads, Pegah Alavi, Randy Garrett, Keysight Technologies) DesignCon 2025, The Road to PCIe 7.0: Advanced Testing Challenges at 64 GBaud PAM4 [https://www.keysight.com/us/en/assets/9925-01141/seminar-materials/KEF-DesignCon-2025-PCIe-Eads-Presentation.pdf]

image-20260607090715624

image-20260427232519514

RLM Measurement Based on Multi-pulse Extraction

image-20260427224538461

image-20260427224752487

reference

Jan Mulder, ISSCC2015 T5: High-Speed Current-Steering DACs

Jihwan Kim, ISSCC2019 F5: Design Techniques for a 112Gbs PAM-4 Transmitter

—, Intel, SNU Summer 2021 [Topic] "A 200Gb/s CMOS Transmitter: Challenges and Overcoming Design Techniques" [https://youtu.be/w3lb_1TwdeE]

—, CICC 2022, ES4-4: Transmitter Design for High-speed Serial Data Communications

Friedel Gerfers, ISSCC2021 T6: Basics of DAC-based Wireline Transmitters

Noman Hai, Synopsys. CICC 2025 Circuit Insights: Basics of Wireline Transmitter Circuits [https://youtu.be/oofViBGlrjM]

—, Synopsys. Design Challenges Of High-Speed Wireline Transmitters [https://semiengineering.com/design-challenges-of-high-speed-wireline-transmitters/]

—, Synopsys. CMOS Circuit Techniques for Wireline Transmitters [https://www.synopsys.com/webinars/wireline-transmitters-part-1.html]

Vivek Gurumoorthy, Marvell, ISSCC 2025 Circuit Insights: Circuits for Optical Communication

Tod Dickson, IBM IEEE San Diego 2020. High-Speed CMOS Serial Transmitters for 56-112Gb/s Electrical Interconnects [https://www.youtube.com/watch?v=g1pcZabsRNc]

B. Razavi, "Design Techniques for High-Speed Wireline Transmitters," in IEEE Open Journal of the Solid-State Circuits Society, vol. 1, pp. 53-66, 2021 [https://www.seas.ucla.edu/brweb/papers/Journals/BROJSSCSep21.pdf]

T. Chan Carusone, T. O. Dickson, S. Palermo, S. Shekhar and M. Mansuri, "Modern Wireline Transceivers," in IEEE Journal of Solid-State Circuits, vol. 61, no. 2, pp. 395-422, Feb. 2026 [https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=11311714]


Yvain Thonnart, CEA-LIST. ISSCC2021 T8: On-Chip Interconnects: Basic Concepts, Designs and Future Opportunities

Mozhgan Mansuri. ISSCC2021 SC3: Clocking, Clock Distribution, and Clock Management in Wireline/Wireless Subsystems

Sam Palermo. High-Performance SERDES Design" Online Course (2025): Current-Mode DAC TX [https://youtu.be/A2VsvCPDWxk]

PCIe® 6.0 Specification: The Interconnect for I/O Needs of the Future PCI-SIG® Educational Webinar Series, [https://pcisig.com/sites/default/files/files/PCIe%206.0%20Webinar_Final_.pdf]

image-20260924012204914

source-degeneration -> +resonator-based CTLE for higher peaking frequency

image-20260925091849782

input network

image-20250706110415914


image-20250611075951974

1
2
3
4
5
>> 10e6/2/pi/400/50

ans =

79.5775

CTLE design target

peaking gain + curve shape

CTLE transfer function

Circuit Insights @ ISSCC2025: Circuits for Wireline Communications - Kevin Zheng [https://youtu.be/8NZl81Dj45M&t=1045]

image-20260328182328792

image-20260328182339762


Why Shunt-peaking or Source Degenerated type Active CTLE? [https://youtu.be/EFMZG-FIWeo]

image-20260519232914332

Shunt Peaking broaden the RC bandwidth

image-20260519233200484

image-20260526210821835

image-20260526210739466

curve shape

PCIe Gen6 Channel and Reference Package S4P Models for Rx Stressed Eye Calibration

image-20251204005909804

Above curve demonstrate that only zero is not enough to compensate channel+pkg loss (>20 dB/decade), peaking or Complex-Conjugate Poles is necessary

image-20251204005738743


S. Shahramian et al., "30.5 A 1.41pJ/b 56Gb/s PAM-4 Wireline Receiver Employing Enhanced Pattern Utilization CDR and Genetic Adaptation Algorithms in 7nm CMOS," 2019 IEEE International Solid-State Circuits Conference - (ISSCC), San Francisco, CA, USA, 2019 [pdf]

image-20251203231733124


P. A. Francese et al., "10.6 continuous-time linear equalization with programmable active-peaking transistor arrays in a 14nm FinFET 2mW/Gb/s 16Gb/s 2-Tap speculative DFE receiver," 2015 IEEE International Solid-State Circuits Conference - (ISSCC) Digest of Technical Papers, San Francisco, CA, USA, 2015 [https://sci-hub.se/10.1109/ISSCC.2015.7062988]

image-20251203232523501


Z. Li, M. Tang, T. Fan and Q. Pan, "A 56-Gb/s PAM4 Receiver Analog Front-End With Fixed Peaking Frequency and Bandwidth in 40-nm CMOS," in IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 68, no. 9, pp. 3058-3062, Sept. 2021 [slides] [paper]

In the active copper cable (ACC) application, it is necessary to give different equalizations at the same frequency according to different cable lengths, Therefore, the AFE with fixed peaking frequency and constant bandwidth is desirable for these applications

image-20251217224711701

Low-Frequency CTLE (LF-CTLE)

S. Parikh et al., "A 32Gb/s wireline receiver with a low-frequency equalizer, CTLE and 2-tap DFE in 28nm CMOS," 2013 IEEE International Solid-State Circuits Conference Digest of Technical Papers, San Francisco, CA, USA, 2013 [https://sci-hub.se/10.1109/ISSCC.2013.6487622]

T. Shibasaki et al., "A 56-Gb/s receiver front-end with a CTLE and 1-tap DFE in 20-nm CMOS," 2014 Symposium on VLSI Circuits Digest of Technical Papers, Honolulu, HI, USA, 2014, pp. 1-2

Yasuo Hidaka Comment #146, #174: Low-Frequency CTLE to support 3m cable w/o FEC [https://www.ieee802.org/3/by/public/Sept15/hidaka_3by_01_0915.pdf]

image-20251217233444843

image-20251217234434659

image-20251217234637716

Equalization Noise Enhancement

Advanced Signal Integrity for High-Speed Digital Designs, S. H. Hall and H. L. Heck, John Wiley & Sons, 2009

CC Chen, Why CTLE? [https://youtu.be/zsuJMqadaKY]

image-20251021211402274

image-20250904235434247

Assuming \(\mathrm{SNR}(f) = \frac{S_x(f)}{S_n(f)}\)


trade-offs between noise amplification and signal equalization

Gm-TIA

H. Kimura et al., "A 28 Gb/s 560 mW Multi-Standard SerDes With Single-Stage Analog Front-End and 14-Tap Decision Feedback Equalizer in 28 nm CMOS," in IEEE Journal of Solid-State Circuits, vol. 49, no. 12, pp. 3091-3103, Dec. 2014 [https://ieeexplore.ieee.org/ielx7/4/6963535/06894632.pdf]

Pisati, et.al., "Sub-250mW 1-to-56Gb/s Continuous-Range PAM-4 42.5dB IL ADC/DAC- Based Transceiver in 7nm FinFET," 2019 IEEE International Solid-State Circuits Conference (ISSCC), 2019 [https://sci-hub.se/10.1109/ISSCC.2019.8662428]

Z. Li, M. Tang, T. Fan and Q. Pan, "A 56-Gb/s PAM4 Receiver Analog Front-End With Fixed Peaking Frequency and Bandwidth in 40-nm CMOS," in IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 68, no. 9, pp. 3058-3062, Sept. 2021 [slides] [paper]

K. Kwon et al., "C12-4 A 212.5Gb/s Pam-4 Receiver With Mutual Inductive Coupled Gm-Tia in 4nm Finfet," 2025 Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits), Kyoto, Japan, 2025, pp. 1-3, doi: 10.23919/VLSITechnologyandCir65189.2025.11074881.

Bae, W. (2019). CMOS Inverter as Analog Circuit: An Overview. Journal of Low Power Electronics and Applications. [pdf]

CTLE, with Gm + TIA structure

image-20260921214634581

image-20250904202636824


Chongyun ZHANG, 2025, "Energy-Efficient CMOS Optical Receiver for Short-Reach Data Center Application,". [slides, paper]

image-20251202222813043

image-20251202222831594

Cherry-Hooper Amplifier

image-20260921235713422

image-20260921235635071

Resonator-Based CTLE

image-20260925162456941

Passive series peaking

D. Pfaff et al., "7.3 A 224Gb/s 3pJ/b 40dB Insertion Loss Transceiver in 3nm FinFET CMOS," 2024 IEEE International Solid-State Circuits Conference (ISSCC), San Francisco, CA, USA, 2024, pp. 128-130, doi: 10.1109/ISSCC49657.2024.10454537.

image-20260922003739424

At DC \[ i_f = g_{md1}v_f \quad i_f = (v_f - v_o)/R_f \qquad \Longrightarrow \qquad \frac{v_o}{v_i} = -\frac{g_{m1}}{g_{md1}} + g_{m1}R_f \] since \(R_f=0\) \[ A_{DC} = -\frac{g_{m1}}{g_{md1}} \]

passive_series_peaking.drawio

\[ \frac{V_o}{V_i}(s) = -g_{m1}R_s\cdot \frac{\frac{1}{LC}}{s^2 + \frac{R_s+R_{LS}}{L}s + \frac{1}{LC}}=-g_{m1}R_s\cdot \frac{\omega_n^2}{s^2+\frac{\omega_n}{Q}s+\omega_n^2} \]

where \[ \boxed{Q=\frac{\omega_n L}{R_s + R_{LS}} \qquad \qquad \omega_n = \frac{1}{\sqrt{LC}}} \]

The \(Q\) is the resonator loaded quality factor

That is \[ A_{DC}=-g_{m1}R_s\qquad\qquad A_{HF}=jg_{m1}R_s\cdot Q \qquad\qquad \frac{A_{HF}}{A_{DC}} = Q \]

where \(A_{HF}\) is the gain at resonance


image-20260922003625379

image-20260922002932197

image-20260922003026369

Q-Shaping (LC-tuned Amplifier)

Y. Krupnik et al., "112 Gb/s PAM4 ADC Based SERDES Receiver for Long-Reach Channels in 10nm Process," 2019 Symposium on VLSI Circuits, Kyoto, Japan, 2019, pp. C266-C267, [https://sci-hub.jp/10.23919/VLSIC.2019.8778136]

—, "112-Gb/s PAM4 ADC-Based SERDES Receiver With Resonant AFE for Long-Reach Channels," in IEEE Journal of Solid-State Circuits, vol. 55, no. 4, pp. 1077-1085, April 2020, [https://sci-hub.jp/10.1109/JSSC.2019.2959511]

S. Kiran et al., "A 56GHz Receiver Analog Front End for 224Gb/s PAM-4 SerDes in 10nm CMOS," 2021 Symposium on VLSI Circuits, Kyoto, Japan, 2021, pp. 1-2, [https://sci-hub.jp/10.23919/VLSICircuits52068.2021.9492471]

Y. Segal et al., "A 1.41pJ/b 224Gb/s PAM-4 SerDes Receiver with 31dB Loss Compensation," 2022 IEEE International Solid-State Circuits Conference (ISSCC), San Francisco, CA, USA, 2022, pp. 114-116, [https://sci-hub.jp/10.1109/ISSCC42614.2022.9731794]

A. Khairi et al., "A 1.41-pJ/b 224-Gb/s PAM4 6-bit ADC-Based SerDes Receiver With Hybrid AFE Capable of Supporting Long Reach Channels," in IEEE Journal of Solid-State Circuits, vol. 58, no. 1, pp. 8-18, Jan. 2023, doi: 10.1109/JSSC.2022.3211475

D. Pfaff et al., "A 224 Gb/s 3 pJ/bit 40 dB Insertion Loss Transceiver in 3-nm FinFET CMOS," in IEEE Journal of Solid-State Circuits, vol. 60, no. 1, pp. 9-22, Jan. 2025, doi: 10.1109/JSSC.2024.3466092

image-20260924003949332

image-20260923233823241

image-20260924002757637

image-20260924011450698


image-20260921234359120

At resonant frequency \[ \boxed{|A_v|= \frac{g_m} {1+\left(\dfrac{g_mR_D}{1+j\omega R_DC_D}\right)} Q_{\mathrm{ind}}\omega L, \qquad \omega=\frac{1}{2\pi\sqrt{L_C L_L}}} \] where \(Q_{\mathrm{ind}}=\frac{\omega L}{R_L}, \qquad R_p=Q_{\mathrm{ind}}^{\,2}R_L=Q_{\mathrm{ind}}\omega L\)

Q-Shaping w/ Parallel RLC

H. Park et al., "7.4 A 112Gb/s DSP-Based PAM-4 Receiver with an LC-Resonator-Based CTLE for >52dB Loss Compensation in 4nm FinFET," 2025 IEEE International Solid-State Circuits Conference (ISSCC), San Francisco, CA, USA, 2025, pp. 142-144, doi: 10.1109/ISSCC49661.2025.10904638.

image-20260925162753669

With Zero-Forcing, inverse of \(h_0 \sim h_1\) only

1
2
3
4
┌ 1    0    0    0 ┐ ┌c0┐   ┌1┐
│ 0.7 1 0 0 │ │c1│ = │0│
│ 0 0.7 1 0 │ │c2│ │0│
└ 0 0 0.7 1 ┘ └c3┘ └0┘
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
% Define the coefficient matrix A
A = [1 0 0 0;
0.7 1 0 0;
0 0.7 1 0;
0 0 0.7 1];

% Define the right-hand side vector b
b = [1;
0;
0;
0];

% Solve the linear system A * c = b for c
% The backslash operator (\) is the recommended way to solve linear systems in MATLAB
c = A \ b;

% Display the result
disp('Vector c:');
disp(c');

% Vector c:
% 1.0000 -0.7000 0.4900 -0.3430

the other method: Treat \(z^{-1}\) as an ordinary variable and divide 1 by \(1+0.7z^{-1}\), working in ascending powers of \(z^{-1}\)

1
2
3
4
5
6
7
8
9
10
11
12
13
                 1  − 0.7z⁻¹ + 0.49z⁻² − 0.343z⁻³ + …      ← quotient = c
┌─────────────────────────────────────────
1 + 0.7z⁻¹ │ 1
│ 1 + 0.7z⁻¹ ← 1 × divisor
│ ───────────
│ − 0.7z⁻¹ ← remainder
│ − 0.7z⁻¹ − 0.49z⁻² ← (−0.7z⁻¹) × divisor
│ ──────────────────
│ + 0.49z⁻²
│ + 0.49z⁻² + 0.343z⁻³ ← (0.49z⁻²) × divisor
│ ───────────────────
│ − 0.343z⁻³
│ …

Use \(1/(1 − r) = 1 + r + r^2 + \dots\) and \(r=-\alpha z^{-1}\) \[ \boxed{\frac{1}{1+\alpha z^{-1}} = 1 - \alpha z^{-1} + \alpha \left( \alpha z^{-2} - \alpha^2 z^{-3} + \dots \right)} \] that is \[ \frac{1}{1+0.7z^{-1}} = 1 - 0.7z^{-1} + 0.7\left( 0.7 z^{-2} - 0.49 z^{-3} + \dots \right) \]

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
% Define the coefficient matrix A
A = [1 0 0 0 0 0;
0.7 1 0 0 0 0;
0.48 0.7 1 0 0 0;
0.35 0.48 0.7 1 0 0;
0.27 0.35 0.48 0.7 1 0;
0.22 0.27 0.35 0.48 0.7 1];

% Define the right-hand side vector b
b = [1;
0;
0;
0;
0;
0];


% Solve the linear system A * c = b for c
% The backslash operator (\) is the recommended way to solve linear systems in MATLAB
c = A \ b;

% Display the result
disp('Vector c:');
disp(c');

% Vector c:
% 1.0000 -0.7000 0.0100 -0.0210 -0.0151 -0.0138

“Inverse of 36 dB LR Channel Response” shows the impulse-response taps of the inverse filter/equalizer.

CTLE1: \(1-\alpha z^{-1}\), without pole

CTLE2: \(1-\alpha z^{-1} + c_1 (\beta z^{-2} - \beta^2 z^{-3} + \dots) = 1 - \alpha z^{-1} + \frac{c_1\beta z^{-2}}{1 + \beta z^{-1}}\), with pole \(-\beta\)

image-20260925180121841

image-20260925184115093

image-20260925184603380

image-20260925184646942

The Summer response is also treated as an impulse-response tap sequence

when the starting sequence is the channel’s pulse response, convolving it with the equalizer’s impulse-response taps gives the equalized pulse-response taps:

\[ p_{\text{out}}[n]=p_{\text{channel}}[n]*h_{\text{EQ}}[n] \]

The distinction is whether the transmitted pulse is already included:

\[ \underbrace{p_{\text{TX}}*h_{\text{channel}}}_{p_{\text{channel}}} *h_{\text{EQ}} =p_{\text{out}} \]

So:

  • Pulse response ∗ impulse response → pulse response.
  • Impulse response ∗ impulse response → combined impulse response.

If CTLE_HF and Summer are both represented by their block impulse responses, their convolution gives the combined impulse response. Including the transmitted pulse gives the pulse response.

shunt peaking

image-20260608215130203

image-20251206000303668

\(\color{red}m=\frac{R^2C}{L}\) is the ratio of the \(R/L\) zero frequency to the original RC pole frequency \(1/RC\), and therefore measures how aggressively the zero compensates the intrinsic RC roll-off.

[Gist link]

image-20260526211605905

image-20260526211624229

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
# Normalize R = C = 1
R = 1.0
C = 1.0

# Frequency sweep
w = np.logspace(-2, 2, 20000)

# Shunt peaking transfer function
#
# R + sL
# H(s) = ---------
# 1 + sRC + s^2 LC
#
# normalized low-frequency gain = 1
#
def H(jw, L):
s = 1j * jw
return (R + s*L) / (R + s*R*R*C + s*s*L*R*C)

# Sweep inductance
Lvals = np.linspace(0.001, 1.5, 300)

image-20260608220405307

image-20260608220440363

[Gist link]

image-20260608224954010

image-20260921235234693


image-20260611225858681

image-20260611225951811

[Gist link]

shunt_chatgpt

image-20260611233156089


ctle_ind

series peaking

image-20260526230433602

image-20260526230805917

T-Coil Peaking

Jri Lee. ISSCC 2009 Tutorial. CMOS Circuit Techniques for High Speed Wireline Transceivers [http://cc.ee.ntu.edu.tw/~jrilee/course/2009_Tutorial_10.pdf]

CC Chen. Why SerDes Needs a Rule of Thumb for T-Coil Design? [https://youtu.be/RIQLYQG2u0A]

Capacitor Splitting + Magnetic Coupling of a transformer

image-20251022234854155

image-20251022235133839

tcoil-tran.drawio


alternative analysis with the below 3 uncoupled inductors model

tcoil-tran-3L.drawio


Three uncoupled inductors model

image-20251126173130097

tcoil-Y.drawio \[\begin{align} V_{P13} &= I_1\cdot sL_1 + I_2\cdot sM = I_1\cdot s(L_1+M) + (I_1-I_2)\cdot s(-M) \\ V_{P23} &= -I_2\cdot sL_2 - I_1\cdot sM = -I_2\cdot s(L_2 + M) + (I_1-I_2)\cdot s(-M) \end{align}\]

The negative inductor \(-M\) can be seen as capacitor \[ -j\omega M = \frac{1}{j}\omega M = \frac{1}{j\omega \frac{1}{\omega^2 M}} \] That is \(C_{-M} = \frac{1}{\omega^2 M} \approx 10 \times C_E\)


image-20260715221444331

T-coil w/ inverted mutual coupling

J. Kim, J. -K. Kim, B. -J. Lee and D. -K. Jeong, "Design Optimization of On-Chip Inductive Peaking Structures for 0.13- μm CMOS 40-Gb/s Transmitter Circuits," in IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 56, no. 12, pp. 2544-2555, Dec. 2009 [https://sci-hub.st/10.1109/TCSI.2009.2023772]

TODO 📅

Triple Resonance

TODO 📅

image-20251206000718234


series resonance \(\omega_\text{res}\)

Assuming, \(I_\text{in}=\cos\omega_r t\) and \(V_\text{out}=g\cos(\omega_r t +\theta)\)

with \(I_\text{in} = C_L\frac{\mathrm{d}V_\text{out}}{\mathrm{d}t}\), yield \(g=\sqrt{\frac{L}{C}}\) and \(\theta=- \frac{\pi}{2}\), i.e. \(V_\text{out} = \sqrt{\frac{L}{C}}\cos(\omega_r t - \frac{\pi}{2})\)

Inductor shorting switches

J. Q. Wang et al., "7.1 A 2.69pJ/b 212Gb/s DSP-Based PAM-4 Transceiver for Optical Direct-Detect Application in 5nm FinFET," 2024 IEEE International Solid-State Circuits Conference (ISSCC), San Francisco, CA, USA, 2024, pp. 123-125 [https://sci-net.xyz/10.1109/ISSCC49657.2024.10454275]

image-20261028203305142

Active Inductor

B. Razavi, "The Active Inductor [A Circuit for All Seasons]," in IEEE Solid-State Circuits Magazine, vol. 12, no. 2, pp. 7-11, Spring 2020 [https://www.seas.ucla.edu/brweb/papers/Journals/BR_SSCM_2_2020.pdf]

activeInd

\[\begin{align} A &= \frac{g_mR_L}{1+(g_{\text{m}_{\text{dio}}}+ g_{\text{ds}_\text{tot}})R_L}\cdot \frac{1+R_pC_Ps}{1+\frac{(1+g_{\text{ds}_{\text{tot}}}R_L)R_PC_P+C_PR_L+R_LC_L}{1+(g_{\text{m}_{\text{dio}}}+g_{\text{ds}_\text{tot}})R_L}s + \frac{R_LC_LR_PC_P}{1+(g_{\text{m}_\text{dio}}+g_{\text{ds}_{\text{tot}}})R_L}s^2} \\ &= \frac{g_mR_L}{1+(g_{\text{m}_{\text{dio}}}+ g_{\text{ds}_{\text{tot}}})R_L}\cdot \frac{R_PC_P}{ \frac{R_LC_LR_PC_P}{1+(g_{\text{m}_{\text{dio}}}+g_{\text{ds}_{\text{tot}}})R_L}}\cdot \frac{1/(R_PC_P)+s}{s^2 + \frac{(1+g_{\text{ds}_{\text{tot}}}R_L)R_PC_P+C_PR_L+R_LC_L}{R_PC_P}s + \frac{1+(g_{\text{m}_{\text{dio}}}+g_{\text{ds}_\text{tot}})R_L}{R_LC_LR_PC_P}} \\ &= A_0 \cdot A(s) \end{align}\]

That is

\[\begin{align} \omega_z &= \frac{1}{R_PC_P} \tag{1} \\ \omega_n &= \sqrt{\frac{1+(g_{\text{m}_{\text{dio}}}+ g_{\text{ds}_\text{tot}})R_L}{R_LC_LR_PC_P}} = \sqrt{\omega_{p0}\omega_z} \\ \zeta & = \frac{(1+g_{\text{ds}_\text{tot}}R_L)R_PC_P+C_PR_L+R_LC_L}{R_PC_P} \frac{1}{2 \omega_n} \end{align}\]

Where \[\begin{align} \omega_{p0} &= \frac{1}{(R_L||\frac{1}{g_{\text{m}_{\text{dio}}}}||\frac{1}{g_{\text{m}_{\text{tot}}}})C_L} \tag{2} \end{align}\]

Here, relate \(\omega_{p0}\) and \(\omega_z\) by coefficient \(\alpha\) \[ \omega_{p0} = \alpha \cdot \omega_z \tag{3} \] This way \[ \omega_n= \sqrt{\alpha}\cdot \omega_z \]

\[ \zeta = \frac{1}{2}(K\sqrt{\alpha}+\frac{1+C_P/C_L}{\sqrt{\alpha}}) \tag{4} \]

where \[ K = \frac{R_L||\frac{1}{g_{\text{m}_{\text{dio}}}}||\frac{1}{g_{\text{m}_{\text{tot}}}}}{R_L||g_\text{ds\_tot}} \]

And \(A(s)\) can be expressed as \[ A(s) = \frac{\frac{s}{\omega_z}+1}{\frac{s^2}{\omega_n^2}+2\frac{\zeta}{\omega_n}s+1} \] It magnitude in dB \[ A_\text{dB} = 10\log\frac{1+(\omega/\omega_z)^2}{1+(\omega/\omega_n)^4+2\omega^2(2\zeta^2-1)/\omega_n^2} \] Substitute \(\omega_n\) with Eq (2), followed is obtained \[ A_\text{dB} = 10\log{\frac{\alpha^2(\omega_z^4 + \omega_z^2\omega^2)}{\alpha^2\omega_z^4+\omega^4+2\alpha\omega_z^2(2\zeta^2-1)\omega^2}} \] peaking frequency \[ \omega_\text{peak} = \omega_z\cdot \sqrt{\sqrt{(\alpha+1)^2 - 4\alpha \zeta^2}-1} \] If \(\zeta=1\) \[ \omega_{A_\text{dB = 0dB} } = \sqrt{1-2/\alpha}\cdot \omega_{p0} \qquad \omega_\text{peak} = \omega_z\sqrt{\alpha-2} \qquad A_\text{dB,peak} = 10\log\frac{\alpha^2}{4(\alpha-1)} \]

Negative Capacitance Circuit

Negative Miller Capacitance

S. Gondi and B. Razavi, "Equalization and Clock and Data Recovery Techniques for 10-Gb/s CMOS Serial-Link Receivers," in IEEE Journal of Solid-State Circuits, vol. 42, no. 9, pp. 1999-2011 [pdf]

Sam Palermo. ECEN620 Lecture 14: Limiting Amplifiers (LAs) [https://people.engr.tamu.edu/spalermo/ecen620/lecture14_ee620_limiting_amps.pdf]

image-20251028221403199

image-20251028232644575 \[ C_{d1} = C_{dd1} + (1+\frac{1}{|A_{gd}|})C_{gd1} \] where \(A_{gd}\lt 0\)

image-20251028232707189

For differential mode input, effective input capacitance \[ C_{in} = C_{gs} +(1+A_{dm}) C_{gd}+\color{red}(1-A_{dm})C_n \] and effective output capacitance \[ C_{out} = C_{dd} + (1+\frac{1}{A_{dm}})C_{gd}+\color{red} (1-\frac{1}{A_{dm}})C_n \] That is \(C_n\) deteriorate the effective output capacitance

For common mode input, effective input capacitance \[ C_{in} = C_{gs} + (1+A_{cm}) C_{gd}+ \color{red}(1+A_{cm})C_n \] and effective output capacitance \[ C_{d1} = C_{dd} + (1+\frac{1}{A_{cm}})C_{gd}+\color{red} (1+\frac{1}{A_{cm}})C_n \] i.e., \(C_n\) deteriorate both effective input capacitance and effective output capacitance, unfortunately


effective input capacitance \(\Pi\) model, which is appropriate for both differential input and common mode input

nmc_pi_in.drawio

Suppose \(C_n=C_{gd}\), effective differential input capacitance is same with effective common-mode input capacitance (\(C_n=\frac{A_{dm}-A_{cm}}{A_{dm}+A_{cm}}C_{gd}\))



XCP with Capacitor

B. Razavi, "The Cross-Coupled Pair - Part III [A Circuit for All Seasons]," IEEE Solid-State Circuits Magazine, Issue. 1, pp. 10-13, Winter 2015. [https://www.seas.ucla.edu/brweb/papers/Journals/BR_Magzine3.pdf]

S. Galal and B. Razavi, "10-Gb/s Limiting Amplifier and Laser/Modulator Driver in 0.18um CMOS Technology,” IEEE Journal of Solid-State Circuits, vol. 38, pp. 2138-2146, Dec. 2003.[https://www.seas.ucla.edu/brweb/papers/Journals/G&RDec03_2.pdf]

A. Sheikholeslami, "Bandwidth Extension [Circuit Intuitions]," in IEEE Solid-State Circuits Magazine, vol. 7, no. 2, pp. 8-11, Spring 2015 [https://www.eecg.utoronto.ca/~ali/papers/mag-spr-15-bandwidth-extention.pdf]

The Cross-Coupled Pair (XCP) can operate as an impedance negator [a.k.a. a negative impedance converter (NIC)]

A common application is to create a negative capacitance that can cancel the positive capacitance seen at a port, thereby improving the speed

image-20240922174319496 \[ I_{NIC} =\frac{V_{im} - V_{ip}}{\frac{2}{g_m}+\frac{1}{sC_c}} = \frac{-2V_{ip}}{\frac{2}{g_m}+\frac{1}{sC_c}} \] Therefore \[ Z_{NIC} = \frac{V_{ip} - V_{im}}{I_{NIC}}=\frac{2V_{ip}}{I_{NIC}} =- \frac{2}{g_m}-\frac{1}{sC_c} \] half-circuit

If \(C_{gd}\) is considered, and apply miller effect. half equivalent circuit is shown as below

nic.drawio


image-20251204223338959

dual-diode based ESD protection

image-20260801014821774

??? which diodes are used in right topology — both pdiode and ndiode are used


image-20260801022107783

reference

T. Chan Carusone, T. O. Dickson, S. Palermo, S. Shekhar and M. Mansuri, "Modern Wireline Transceivers," in IEEE Journal of Solid-State Circuits, vol. 61, no. 2, pp. 395-422, Feb. 2026 [https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=11311714]

Miguel Gandara. CICC2025 Circuits Insights: Wireline Receiver Circuits [https://youtu.be/X4JTuh2Gdzg]

Elad Alon, ISSCC 2014, "T6: Analog Front-End Design for Gb/s Wireline Receivers"

Byungsub Kim, ISSCC 2022, "T11: Basics of Equalization Techniques: Channels, Equalization, and Circuits"

Gain Kim, 2023. Equalization, Architecture, and Circuit Design for High-Speed Serial Link Receiver [www.theise.org/...]

Nhat Nguyen and Masum Hossain, ISSCC 2021 Forum F6.7: 112Gb/s-and-Beyond Long-Reach and Short-Reach Electrical Interfaces

Jihwan Kim, Intel, ISSCC 2023 Forum F1.5: Circuit Designs for 200+Gb/s Electrical Transceivers

Ariel Cohen, Intel, ISSCC 2024 Forum F6.3: Beyond 200Gbps Electrical transceivers – Circuit Architecture, Design Implementation and Silicon Results

Heng Zhang, Broadcom, ISSCC 2025 Forum F4.2: High-speed ADCs for 100Gbps+ Wireline Transceivers

E-Hung Chen, MTK, ISSCC 2026 Forum F2.3: State-of-the-Art 200+ Gb/s Electrical and Optical Interconnects


J. Kim et al., "A 112Gb/s PAM-4 transmitter with 3-Tap FFE in 10nm CMOS," 2018 IEEE International Solid-State Circuits Conference - (ISSCC), San Francisco, CA, USA, 2018 [paper]

S. Shekhar, J. S. Walling and D. J. Allstot, "Bandwidth Extension Techniques for CMOS Amplifiers," in IEEE Journal of Solid-State Circuits, vol. 41, no. 11, pp. 2424-2439, Nov. 2006 [pdf]

David J. Allstot Bandwidth Extension Techniques for CMOS Amplifiers [https://ewh.ieee.org/r5/denver/sscs/Presentations/2007_08_Allstot.pdf]

S. S. Mohan, M. D. M. Hershenson, S. P. Boyd and T. H. Lee, "Bandwidth extension in CMOS with optimized on-chip inductors," in IEEE Journal of Solid-State Circuits, vol. 35, no. 3, pp. 346-355, March 2000 [http://smirc.stanford.edu/papers/JSSC00MAR-mohan.pdf]

J. Paramesh and D. J. Allstot, "Analysis of the Bridged T-Coil Circuit Using the Extra-Element Theorem," in IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 53, no. 12, pp. 1408-1412, Dec. 2006 [https://sci-hub.st/10.1109/TCSII.2006.885971]

S. C. D. Roy, "Comments on "Analysis of the Bridged T-coil Circuit Using the Extra-Element Theorem," in IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 54, no. 8, pp. 673-674, Aug. 2007 [https://sci-hub.st/10.1109/TCSII.2007.899834]

B. Razavi, "The Bridged T-Coil [A Circuit for All Seasons]," IEEE Solid-State Circuits Magazine, Volume. 7, Issue. 40, pp. 10-13, Fall 2015 [https://www.seas.ucla.edu/brweb/papers/Journals/BRFall15TCoil.pdf]

—, "The Design of Broadband I/O Circuits [The Analog Mind]," IEEE Solid-State Circuits Magazine, Volume. 13, Issue. 2, pp. 6-15, Spring 2021 [http://www.seas.ucla.edu/brweb/papers/Journals/BR_SSCM_2_2021.pdf]

Deog-Kyoon Jeong. Topics in IC Design: T-Coil [pdf]

P. Heydari, "Neutralization Techniques for High-Frequency Amplifiers: An Overview," in IEEE Solid-State Circuits Magazine, vol. 9, no. 4, pp. 82-89, Fall 2017 [https://sci-hub.ru/10.1109/MSSC.2017.2745858]

—, "Evolution of Broadband Amplifier Design: From Single-Stage to Distributed Topology," in IEEE Microwave Magazine, vol. 24, no. 9, pp. 18-29, Sept. 2023

Cowan G. Mixed-Signal CMOS for Wireline Communication: Transistor-Level and System-Level Design Considerations. Cambridge University Press; 2024

Starič, Peter and Erik Margan. Wideband amplifiers. (2006) [pdf]

Bob Ross. IBIS Summit [T-Coils and Bridged-T Networks], [T-Coil Topics]

Walling, Jeffrey & Shekhar, Sudip & Allstot, David. (2008). Wideband CMOS Amplifier Design: Time-Domain Considerations. Circuits and Systems I: Regular Papers, IEEE Transactions on. 55. 1781 - 1793. [pdf]

A. A. Abidi, "The T-Coil Circuit Demystified," in IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 72, no. 9, pp. 4469-4480, Sept. 2025

S. Lin, D. Huang and S. Wong, "Pi Coil: A New Element for Bandwidth Extension," in IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 56, no. 6, pp. 454-458, June 2009

M. Kossel et al., "A T-Coil-Enhanced 8.5 Gb/s High-Swing SST Transmitter in 65 nm Bulk CMOS With <−16 dB Return Loss Over 10 GHz Bandwidth," in IEEE Journal of Solid-State Circuits, vol. 43, no. 12, pp. 2905-2920, Dec. 2008 [https://web.mit.edu/magic/Public/papers/04684644.pdf]


S. Galal and B. Razavi, "Broadband ESD protection circuits in CMOS technology," in IEEE Journal of Solid-State Circuits, vol. 38, no. 12, pp. 2334-2340, Dec. 2003, [https://sci-hub.jp/10.1109/JSSC.2003.818568]

M. Ker and Y. Hsiao, "On-Chip ESD Protection Strategies for RF Circuits in CMOS Technology," 2006 8th International Conference on Solid-State and Integrated Circuit Technology Proceedings, 2006, pp. 1680-1683 [https://sci-hub.jp/10.1109/ICSICT.2006.306371]

M. Ker, C. Lin and Y. Hsiao, "Overview on ESD Protection Designs of Low-Parasitic Capacitance for RF ICs in CMOS Technologies," in IEEE Transactions on Device and Materials Reliability, vol. 11, no. 2, pp. 207-218, June 2011 [https://sci-hub.jp/10.1109/TDMR.2011.2106129]

Kosnac, Stefan (2021) Analysis of On-Chip Inductors and Arithmetic Circuits in the Context of High Performance Computing [https://archiv.ub.uni-heidelberg.de/volltextserver/30559/1/Dissertation_Stefan_Kosnac.pdf]

Chapter 4.5. High Frequency Passive Devices [https://www.cambridge.org/il/files/7713/6698/2369/HFIC_chapter_4_passives.pdf]


K. Yadav, P. -H. Hsieh and A. Chan Carusone, "Linearity Analysis of Source-Degenerated Differential Pairs for Wireline Applications," in IEEE Open Journal of Circuits and Systems [https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=10769573]

Minsoo Choi et al., "An Approximate Closed-Form Channel Model for Diverse Interconnect Applications," IEEE Transactions on Circuits and Systems-I: Regular Papers, vol. 61, no. 10, pp. 3034-3043, Oct. 2014. [https://sci-hub.jp/10.1109/TCSI.2014.2327275]

0%