
Frequency Domain Model


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

Direct
Interleaver

similar to increase the resolution of the flash ADC with
more parallel comparators
De-multiplexing
Interleaver

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

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

Interleaver
Model

Interleaving Errors


Offset Mismatch
Errors

Gain Mismatch
Errors



Timing Mismatch
Errors

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



Bandwidth Mismatch
Errors


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

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

reset
mechanism to the latches for nominal order

a clock generator
for an eight-channel ADC

| \(\phi_1\) |
\(\text{CK}\) |
\(I_1\) |
\(I_2\) |
| \(\phi_2\) |
\(\overline{\text{CK}}\) |
\(Q_1\) |
\(I_3\) |
| \(\phi_3\) |
\(\text{CK}\) |
\(\overline{I_1}\) |
\(Q_2\) |
| \(\phi_4\) |
\(\overline{\text{CK}}\) |
\(\overline{Q_1}\) |
\(Q_3\) |
| \(\phi_5\) |
\(\text{CK}\) |
\(I_1\) |
\(\overline{I_2}\) |
| \(\phi_6\) |
\(\overline{\text{CK}}\) |
\(Q_1\) |
\(\overline{I_3}\) |
| \(\phi_7\) |
\(\text{CK}\) |
\(\overline{I_1}\) |
\(\overline{Q_2}\) |
| \(\phi_8\) |
\(\overline{\text{CK}}\) |
\(\overline{Q_1}\) |
\(\overline{Q_3}\) |
\(\text{CK}\), \(I/Q_1\) 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]


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]


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


Ensure nominal order by
masking some edges


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






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.


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]








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


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]

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
| 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


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

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.
| 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 |
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.

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 \(f_s/2\).


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.

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
| 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.

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:

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
| 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

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

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

\[
D_n(kT) = (G_n R(kT) V(kT) + O_n)R(kT)= C_n V(kT) + R(kT)O_n
\]
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 \(V_{CM}\) \[
\sigma^{VCM}_{DNL,\max}
=
\sqrt{2^{n-1}-1}\,\frac{\sigma_u}{C_u},
\qquad
\sigma^{VCM}_{INL,\max}
=
\frac{\sqrt{2^{n-1}}}{2}\,\frac{\sigma_u}{C_u}
\qquad
\text{at } k=2^{n-2},\,3\cdot 2^{n-2}
\]

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]