CTLE Adaptation

Sam Palermo has a nice overview of 3 different CTLE adaptation schemes
http://www.ece.tamu.edu/~spalermo/ecen689/lecture8_ee720_rx_adaptive_eq.pdf

Which lead me to the following paper:

F. Gerfers, G. W. den Besten, P. V. Petkov, J. E. Conder and A. J. Koellmann, "A 0.2–2 Gb/s 6x OSR Receiver Using a Digitally Self-Adaptive Equalizer," in IEEE Journal of Solid-State Circuits, vol. 43, no. 6, pp. 1436-1448, June 2008.

My task is to understand this enough to implement it in my model.

• The CTLE's objective is to reduce the ISI observed in the post-cursor positions of the pulse response.
• OSR = over sampling rate
• "Both, the data eye opening and the distribution of the data edges over the transition period depend on random and deterministic effects introduced by the TX, channel and RX. The deterministic portion is dominated by limited channel bandwidth introducing ISI whereas the random parts are caused by noise sources in the TX and RX front-ends. As the impact of the random effects are averaged over time, the width and the shape of the data edge distribution can be reliably used to calibrate the equalizer [20].
• "The edge information extracted from the data samples are used to determine the histogram of the signal transitions as shown in Fig. 10. The histogram analysis is implemented by using six XOR gates, each one comparing two successive sample phases to detect the occurrence of data transitions between these phases. The number of detected edges between all pairs of two consecutive sampling phases is accumulated over a selectable number of signal transitions using six variable transition counters X_i each Y_i 12-bit depth averaging out random effects. As a result, the width of the transition histogram is given by the number of transition counters X_i with non-zero content Y_i != 0 while the shape is evaluated by the standard deviation of the histogram.

• "the proposed equalizer tuning targets to maximize the eye-opening and minimizing at the same time the standard deviation of the edge distribution (narrow the histogram shape).

• A simple (straightforward) estimation defines the maximum bin of the histogram max(X_i) as the mean value (mu) of the histogram and calculates the standard deviation (sigma) by applying a quadratically binary weighted calculation
• However, it is preferable to have continuous adaptation to handle issues of changing channel characteristics due to e.g. moving the cable or temperature variations. This is the objective for the next receiver generation by adding an extra equalizer tuning bit and a glitch free equalizer tuning scheme which allows modification of equalizer settings on-the-fly without temporary data loss.

I'm disappointed! Their adaptation scheme is really an exhaustive search and pick the best approach.  What I really need is something that gives an indication if the signal is under or over equalized and move accordingly.

Behavioral RX Model Correlation with Physical Devices, and Accuracy Improvement - DesignCon 2015

Behavioral RX Model Correlation with Physical Devices, and Accuracy Improvement - DesignCon 2015, Altera

I am interested in this paper because I think that they will give clues on how to extract the CTLE response of a receiver from measurement somehow.

Well, I read most of the paper but I really didn't understand how they measured the CTLE high frequency gain with only their receiver BIST reporting BER.  I wonder if they had an eye scan capability and you put through clock patterns of various frequency and amplitude if you could easily extract the CTLE response.

RX IBIS-AMI Model Silicon Correlation Metrics and Model Development Methodology - DesignCon 2017

RX IBIS-AMI Model Silicon Correlation Metrics and Model Development Methodology - DesignCon 2017, Masashi Shimanouchi of Intel Corporation

First off, I really enjoyed the presentation.  The presenter had these stick figures having conversations between the AMI model developer and the user that made everyone laugh.  They went something like this, "Here is the AMI model", "how accurate is it?", "It's really good", "give me a number", "90% accurate", "so it will correlate to hardware 90% of the time?", "something like that".

He mentioned that in a previous work that they were able to reliably extract the CTLE response of a receiver from measurements which I am very interested in.  I need to look up their previous 2 DesignCon papers for more information.  These guys all came from the Intel acquisition of Altera.

Notes on the paper:

• "Constraints on the observables demand different metrics for pre-silicon model and post-silicon model. We found BER and optimum equalizer gain the most meaningful metrics to quantitatively evaluate the level of silicon correlation of RX IBIS-AMI model.
• They make extensive use of the jitter tolerance tests (JTOL).  I need to figure out a good way of performing this test with QCD.
• We should add confidence intervals to the QCD TD BER bathtub plots
• They focus on correlating the worst case and typical corners.  While this is nice way to reduce the correlation scope, it might be interesting to reserve the best case corner for a validation set.  Where the tuning of the AMI model occurs at the worst case and typical corners and the resulting model is compared to the best case corners to see how general the AMI model is.
• The study was really limited by the lab measurements of BER.  They ran the link for 3e6 bits, 5 minute at 10 Gbps and if they observed no error then the BER was determined to be 1e-12 with 95% confidence.  This resulted in granular measurements which are difficult to correlate to nearly continuous simulation results.  I bet next time they are going to get finer lab measurements so their correlation will be better.

Polynomial Chaos Expansion - DesignCon 2017

"Exploring Efficient Variability-Aware Analysis Method for High-Speed Digital Link Design Using PCE" presented by Jan B. Preibisch1 at DesignCon 2017

I first met Jan at the SPI2016 conference in Turin Italy last year and am very interested in his work as it meets several of the deficiencies of Design of Experiments.

• Polynomial chaos expansion "is a spectral approach which projects the stochastic variables onto an orthogonal polynomial basis.  This approach converges orders of magnitude faster than MC and the expansion coefficients inherently comprise stochastic measures like mean and variance."
• SGM=stochastic Galerkin matching
• ST = stochastic testing
• "Monte Carlo simulation is a gold standard for statistical analysis methods where the parameter values are randomly selected based on their probability distributions"
• " Mathematically, the stochastic input parameter is represented as a linear function of the unitless stochastic variable. The general idea of PCE is to represent the desired output parameter as a function of the same stochastic variable. As the word polynomial in PCE suggests, the function is assumed to be a polynomial. Orthogonal polynomials exhibit extraordinary convergence properties and are chosen to construct a proper mathematical framework.

This is an interesting paper but I don't have any plans to implement it anytime soon so I am going to leave the math for another day.