Multiple-Population Moment Estimation: Exploiting Inter-Population Correlation for Efficient Moment Estimation in Analog/Mixed-Signal Validation

Chenjie Gu, Manzil Zaheer, Xin Li

TL;DR

This paper introduces Multiple-Population Moment Estimation (MPME), a novel Bayesian approach that significantly boosts the accuracy of estimating circuit performance moments (like mean and variance) when dealing with very limited data. By exploiting correlations between data from different circuit configurations or "corners," MPME achieves up to a 2x reduction in estimation error, directly translating to reduced validation time and cost for complex analog/mixed-signal circuits.

What Problem Does This Paper Address?

Validating complex analog and mixed-signal circuits is crucial for ensuring product quality and guiding design optimization. A key part of this validation involves estimating the statistical distribution of performance metrics (like signal delay or power consumption) to calculate metrics like Probability of Failure (PoF) and parametric yield. However, for these circuits, obtaining data through simulation or measurement is extremely time-consuming and expensive. This often results in a severe "small-sample-size" problem, where the limited data available can lead to inaccurate and unreliable estimates. This issue is further compounded because validation needs to be performed across numerous "corners" (e.g., different process variations, voltages, temperatures) and configurations (e.g., different board setups), meaning each individual corner/configuration might have very few data points. Traditional methods struggle to provide accurate estimates under these constraints.

Methods

The core innovation is the Multiple-Population Moment Estimation (MPME) method. It frames the problem within a hierarchical Bayesian framework. The key idea is to treat data from different populations (defined by specific corners/configurations) not as independent, but as correlated.

MPME employs a two-step process:

  1. Learning the Prior Distribution: A generative graphical model is used to represent the relationships between the moments (mean and variance) of different populations. This model allows the algorithm to learn a joint prior distribution over these moments from the data collected across all populations. This is done using Maximum Likelihood Estimation (MLE) to find the hyperparameters of the prior. Two specific prior distributions are explored:
  2. Uniform Prior (UNI): Assumes means and variances fall within specified bounds.
  3. Normal-Inverse-Chi-Squared Prior (NIX): A conjugate prior that allows for efficient closed-form solutions and can be interpreted as adding "fake" data samples to boost effective sample size.
  4. MAP Estimation: The learned prior distribution is then used to perform Maximum A Posteriori (MAP) estimation for the moments of each individual population. This Bayesian approach effectively "shrinks" estimates towards the learned prior, leveraging information from other populations to improve accuracy, especially when individual sample sizes are small.

Datasets

The paper evaluates MPME using:

  • Synthetic Datasets: Generated to systematically study the impact of sample size (N) and number of populations (P) on estimation accuracy.
  • Post-Silicon Measurements: Data from a commercial high-speed I/O link, specifically measuring time margins (which are related to Bit-Error-Ratio or BER). This real-world dataset demonstrates the practical applicability of MPME.

Key Insights and Ideas

  • Inter-Population Correlation is Key: The most significant insight is that data from different circuit corners/configurations are not independent but correlated. Exploiting this correlation is the fundamental mechanism by which MPME improves accuracy.
  • Bayesian Shrinkage for Small Samples: MPME acts as a form of "Bayesian shrinkage," where estimates for individual populations are pulled towards a common learned distribution. This is highly effective when individual sample sizes are small, preventing estimates from being overly influenced by noisy data points.
  • Prior Choice Matters (and is Learned): The choice of prior distribution (UNI vs. NIX) influences the estimation. NIX often provides better overall accuracy by pulling estimates towards a learned global mean, while UNI can be more robust for populations with extreme values by imposing bounds. Crucially, the hyperparameters of these priors are learned from the data itself, making the method adaptive.
  • Effective Sample Size Increase: The NIX prior, in particular, can be interpreted as effectively increasing the sample size for each population, leading to more confident estimates.
  • Beyond Mean and Variance: While the paper focuses on the first two moments, the underlying graphical model framework is extensible to other moments or even full density estimation for non-Gaussian distributions.

Key Results

  • Significant Accuracy Improvement: MPME consistently outperforms traditional sample mean and variance estimators, achieving an average error reduction of up to 2x in experiments.
  • Performance with Small Samples: The benefits of MPME are most pronounced when the number of samples per population (N) is very small (e.g., N=5).
  • Scalability with Populations: As the number of populations (P) increases, MPME's error decreases roughly as 1/√P, demonstrating its ability to leverage more correlated data.
  • Reduced Validation Time: The accuracy improvement directly translates to potentially significant reductions in validation time and cost, as fewer samples are needed to achieve a desired confidence level. For example, in the I/O link validation, MPME-NIX required only about 50% of the samples compared to traditional methods for similar accuracy.

Limitations and Future Work

  • Convexity of Optimization: The optimization problems in MPME (learning prior hyperparameters and MAP estimation) are not guaranteed to be convex, meaning local optima might be found. However, the authors suggest that initial guesses from the data help mitigate this.
  • Potential for Bias: In extreme, pathological cases (e.g., many populations with very different characteristics), MPME might introduce bias for specific populations by shrinking their estimates too aggressively. The authors note that such cases are unlikely in practice and outliers can be handled.
  • Higher-Order Moments: While theoretically extensible, estimating higher-order moments (like skewness or kurtosis) with MPME is challenging due to the difficulty in defining the probability distribution from moments and potential for large estimation errors with small samples.
  • Future Work: The paper suggests extending MPME to non-Gaussian distributions and higher-order moments, and exploring more complex prior distributions for even greater flexibility.

Summary updated

Abstract

Moment estimation is an important problem during circuit validation, in both pre-Silicon and post-Silicon stages. From the estimated moments, the probability of failure and parametric yield can be estimated at each circuit configuration and corner, and these metrics are used for design optimization and making product qualification decisions. The problem is especially difficult if only a very small sample size is allowed for measurement or simulation, as is the case for complex analog/mixed-signal circuits. In this paper, we propose an efficient moment estimation method, called Multiple-Population Moment Estimation (MPME), that significantly improves estimation accuracy under small sample size. The key idea is to leverage the data collected under different corners/configurations to improve the accuracy of moment estimation at each individual corner/configuration. Mathematically, we employ the hierarchical Bayesian framework to exploit the underlying correlation in the data. We apply the proposed method to several datasets including post-silicon measurements of a commercial high-speed I/O link, and demonstrate an average error reduction of up to 2$\times$, which can be equivalently translated to significant reduction of validation time and cost.
BibTeX
@article{gu2014multiple,
  author = {Chenjie Gu and Manzil Zaheer and Xin Li},
  title = {{Multiple-Population Moment Estimation: Exploiting Inter-Population Correlation for Efficient Moment Estimation in Analog/Mixed-Signal Validation}},
  journal = {IEEE Trans. Comput. Aided Des. Integr. Circuits Syst.},
  year = {2014},
  url = {https://openreview.net/forum?id=MsCE6I4gJ2},
}
Figure