Spectral Prediction for Improved Loopback Test of Embedded Mixedsignal Circuits

1
Spectral Prediction for Improved Loopback Test
of Embedded Mixed-signal Circuits

• Hongjoong shin, Byoungho Kim
• and Jacob A. Abraham
• Computer Engineering Research Center
• The University of Texas at Austin

2
Outline

• Component-level tests Vs. System-level tests
• Fault masking
• Prior work
• Dual Loopback tests - Proposed method
• Spectral Prediction
• Simulation results
• Design Considerations of Built-in Test Circuit
• Conclusions

3
Component-level test

• Component-level test
• Isolated tests for each analog block
• Sequential tests
• Advantage
• High fault coverage
• However,
• Test time overhead
• More difficult in Mixed-signal SOCs
• Due to limited I/O, limited resources of ATE

4
System-level test

• System-level test
• Tests several blocks at the same time
• Parameters such as Bit-Error-Rate and
  Signal-to-Noise and Distortion Ratio (SINAD) are
  measured
• Advantages
• Reduces test time
• Avoids performance degradation due to test point
  insertion
• Self-tested
• Examples
• DAC-ADC loopback
• RX/TX loopback

5
Issues of system-level test

• Fault masking can be caused by interaction
  between non-functionally related blocks
• Fault Escape
• Overqualified component masks a fault of faulty
  component in loopback path
• Misclassifications!!
• Yield Loss
• Marginally fault-free components results in
  faulty loopback response

6
Fault masking simulation I

• 2200 DUT ensembles were generated by statistical
  variations in noise and distortion
• 200 DUTs were simulated in loopback and normal
  mode to set the limit of loopback response
• Loopback SINAD of remaining DUTs was used for
  classification

14bits of resolution Operated at 100kHz
7
Fault masking simulation II
LP, AF

• Actual Pass 92
• Loopback Pass 90
• Misclassifications
• Loopback Pass, but Actual Fail 8.2
• Actual Pass, but Loopback Fail 2.3
• Misclassification strongly depends on tolerance
  bands of ADC/DAC and distribution

LF, AP
2
LF, AF
8
6
84
Loopback Pass (LP) Actual Pass (AP)
8
Fault masking Why?

• Based on sine-wave modeling

ß1, ß2, ß3, nß(t)
ylb(t)
ydac(t)
a1, a2, a3, na(t)
Cannot determine unique solution for ai and ßi
Loopback response does not directly represent
performances of ADC and DAC
9
Prior work

• Hybrid BIST proposed by Ohletz
• DUT is reconfigured into DAC-ADC loopback
• Only pass/fail of test is possible
• Loopback with internal node monitoring
• Proposed by Yu
• Characterize Sigma-delta ADC/DAC
• Only noise parameter splits

10
Dual loopback response -Proposed method

• Dual loopback responses are generated using
  built-in test circuit
• Normal loopback response
• Filtered loopback response
• Filtered loopback
• Scaled output of filter inputs to ADC

11
Dual loopback response II

• Assume A ß1 a1 1
• Unity gain of ADC, DAC and input signal
• Then, dual loopback responses are
• Possible to determine unique solution for ai, ßi
  and noise

12
Dual loopback response III

• However
• prediction error can be caused by
• Variations of gain A, ß1 and a1
• It may lead to convergence problem
• Uses statistical prediction model
• Prediction model is built using Multivariate
  Adaptive Regression Splines (MARS, Friedman 1991)
• Spectral predictors are used as input variables

Solution space
13
Prediction using MARS

• Introduced by Variyam, 1998
• Popular in signature-based tests
• Used to find mapping function btw output response
  (signature) and specifications
• Mapping function is built using MARS
• Predicted performance parameters are used for DUT
  classification
• Signature is measured by transient sampling

14
Issues in transient predictors

• Transient prediction requires
• Measurements to be either synchronized or
  precisely triggered to provide stationary
  sampling
• Phase error to be characterized prior to its
  application to prediction model
• Requires sufficient predictor variables to
  increase the prediction accuracy
• Solution
• Uses Spectral predictors

15
Modeling with Spectral predictors

• Frequency response is used as predictor variables
  through spectral sampling
• Frequency response is obtained using Discrete
  Fourier Transform (DFT)
• Spectral powers at frequencies of interest are
  used as predictor variables for MARS modeling
• DC, fundamental, 2nd harmonic, 3rd harmonic,
  Average noise.. , etc

16
Advantages of Spectral predictors
17
Example DAC-ADC loopback

• DAC-ADC loopback was modeled and simulated using
  MATLAB
• Performance parameters considered were SINAD,
  SNR, THD and Gain

• 2200 DUT ensembles were generated with normal
  distribution in value of gain, noise and
  distortion
• Non-idealities of filter are included

18
Results

• Actual SINAD Vs. Predicted SINAD

19
Results

• Summary

20
Design considerations

• Filter Specifications
• Center Frequency
• Fundamental
• Gain
• Unity
• Roll-off Slope (Nth-order)
• Second-order
• Quality Factor (Q)
• Noise

Second-Order Sallen-Key Filter
21
Design considerations - Q

• Prediction Errors Vs. different Q values

22
Design considerations - Q

• Prediction errors are relatively insensitive to
  the value of Q
• 0.05dB improvements for 10x
• Some thoughts on Q
• Higher Q induces in-band noise
• Noise of bandpass filter Vs. Prediction Error

23
Design considerations - Noise

• Prediction Errors Vs. different ratios filter
  noise to DUT noise

24
Design considerations - Noise

• Lower filter noise provides more accurate
  prediction
• Linearly increases
• Prediction errors are still within 1.2dB for 200
  noise ratio
• Go back to Quality factor
• Higher Q Higher in-band noise
• Use lower Q

25
Conclusions

• Performance parameters such as SINAD, SNR, THD
  and Gain can be predicted
• Can be extended to predict SFDR and
  intermodulation distortion
• Prediction errors were within 0.6dB
• 1/10 bit error
• This technique can be applied to general types of
  mixed-signal circuits test using loopback

26
Thank you!! Any Questions?