This document presents a perfect data reconstruction algorithm for interleaved ADCs. It begins with an overview of the purpose and challenges of time-interleaved ADCs, namely that individual ADCs and analog sections can have different offsets, gains and phases. It then discusses how to represent the relationship between input frequencies and the mutually aliased frequencies captured by each ADC in a matrix equation. The document outlines how to measure the complex gain coefficients for this matrix and solve the set of linear equations to reconstruct the input signal spectrum. It validates this approach on hardware data, achieving a 54dB SNR and -78dBc spurs, outperforming individual ADCs. The algorithm provides the basis for high-performance instruments but has limitations related
Perfect data reconstruction algorithm of interleaved adc
1. Perfect data reconstructionPerfect data reconstruction
algorithm of interleaved ADCalgorithm of interleaved ADC
Dr. Fang Xu
Teradyne, Inc. Boston, MA, U.S.A.
3. PurposePurpose
Test instruments are built with available parts
Instrument development time is long
Instruments are designed for testing future products
Performance gap needs be solved by design
Year
Performance
(frequency, bits)
Instrument
architecture
reduces
Performance gap
State of art device performance
Future
Product
Instrument
Design-in
4. Time Interleaved ADC’sTime Interleaved ADC’s
Capture of a continuous time domain waveform
ADC7
ADC6
ADC5
ADC4
ADC3
ADC2
ADC1
ADC0
Clock generation
Interleaved
samples
5. Time Interleaved Real ADC’sTime Interleaved Real ADC’s
ADC’s and analog sections have different
offset, gain and phase
Gain and phase vary with frequency
Up-to 20 dB measured for gain !
Samples are not uniformly distributed
Need advanced algorithm to reconstruct signal
Relative gain/phase (timing) error vs. 1st
ADC @199.99 MHz
5 dB/div 50 ps/div
6. Time Interleaved Real ADC’sTime Interleaved Real ADC’s
ADC’s and analog sections have different
offset, gain and phase
Gain and phase vary with frequency
Up-to 20 dB measured for gain !
Samples are not uniformly distributed
Need advanced algorithm to reconstruct signal
ADC7
ADC6
ADC5
ADC4
ADC3
ADC2
ADC1
ADC0
Input
Clock generation
Data
correction
reconstruction
7. FFT of Capture Before CorrectionFFT of Capture Before Correction
H2
offset
gain/phase
-120
-80
-40
Fi = 199.990200 MHz,
Fs = 1.494220800
Gsamples/s
SNR=20 dBc,
Non harmonic spur=-25
dBc
100 200 300 400 500 600 700
9. Gain Discrepancy ArtifactsGain Discrepancy Artifacts
0
Repetitive amplitude modulation
Spur at ± input tone to integer Fs
Need advanced algorithm
100 200 300 400 500 600 700
H2
offset
-120
-80
-40
gain/phase
10. Phase/Timing Discrepancy ArtifactsPhase/Timing Discrepancy Artifacts
0
Repetitive phase modulation
Spur at ± input tone from integer Fs
Need advanced algorithm
H2
offset
-120
-80
-40
gain/phase
100 200 300 400 500 600 700
11. Sampling and Aliasing at FsSampling and Aliasing at Fs
Aliased in frequency domain without Hermitian
symmetry
Redundant information with Hermitian symmetry
Alias
Alias
12. Family of Mutually Aliased FrequenciesFamily of Mutually Aliased Frequencies
Repetitive amplitude/phase modulation
Spur at ± input tone from integer Fs
That is a subset of whole spectrum
-40
gain/phase
100 200 300 400 500 600 700
We call this subset of frequencies including that of signal
A family of mutually aliased frequencies (FMAF)
Frequencies number equals the number of ADCs
Vector notation: iNMiMNiNkNikNiNi XXXXXX +−−−++− )12/(
*
)2/(
*
)(
*
,,,,,
-20
13. Frequency Domain ReconstructionFrequency Domain Reconstruction
Fs
Input signal spectrum
to be reconstructed
ADC7
ADC6
ADC5
ADC4
ADC3
ADC2
ADC1
ADC0
Clock generation
Fs/2
Spectrum at output of each ADC
Matrix of linear system FMAF
Orthogonal components outside FMAFPorous matrix (lot of 0)
Sampling with Hermitian symmetrySmall matrix for each FMAF
15. Unknowns and Knowns in EquationUnknowns and Knowns in Equation
Fs
Component at frequency i
=
+−
+
−
−+
−
iNM
ikN
i
iN
iNkN
iMN
R
X
X
X
X
X
X
)12/(
*
*
)(
*
)2/(
ˆ
X
=
− iM
im
i
R
X
X
X
,1
,
,0
~
.
~
.
~
~
X
Unknown:
All frequency
components within
a FMAF
Captured
data of all
converters
Fs/2
Captured data
of converter m
at frequency i
iX
imX ,
~
17. ADC7 FFT
Fs
ADC6 FFT
ADC5 FFT
ADC4 FFT
ADC3 FFT
ADC2 FFT
ADC1 FFT
ADC0 FFT
Input
N/2 times
MxM linear
equations
Order of data
Frequency Domain ReconstructionFrequency Domain Reconstruction
Solving linear equations for each FMAF
Reorder data according to Hermitian symmetry
RRR XHX
~ˆ 1−
=
18. -120
-100
-80
-60
-40
-20
Magnitude(dBFS)
Correction Result of Captured SignalCorrection Result of Captured Signal
Fi = 199.9902 MHz, Fs = 1.4942208
Gmples/s
Before correction
SNR= 20dBc, Non harmonic spur= -25dBc
After correction
SNR= 54dBc, Non harmonic spur= -78dBc
100 200 300 400 500 600 700
19. DiscussionsDiscussions
Performance
54dBc SNR @750MHZ BW = 142dBc/Hz
limited by signal generator
-78dBc Spur –20dB dispersion better than
SFDR of ADC
Hardware stability limitation
Ex: A 0.1% converter gain change will limit
performance level to about -60dB
This does not cover non-linear distortion
Application limitation
DFT can only start when entire segment of
waveform has been captured
This method is a better fit for applications
which do not need real time capture
20. ConclusionsConclusions
Solution based on general model of ADC
Gain and phase are functions of frequency
Complete mathematical resolution
Validation by data captured on hardware
Results exceed expectation
Base of high-performance instruments
21. Perfect data reconstructionPerfect data reconstruction
algorithm of interleaved ADCalgorithm of interleaved ADC
Questions and Answers
? And !