Enhanced metamodeling techniques for high-dimensional IC design estimation problems

Andrew B. Kahng, Bill Lin, Siddhartha NathDATE 2013 · DOI

TL;DR

This paper investigates the effectiveness of various metamodeling techniques for estimating key IC design metrics in high-dimensional spaces. They find that adaptive sampling significantly improves efficiency and accuracy, and that the best metamodeling approach (tree-based vs. Gaussian process-based) depends heavily on the dimensionality of the problem. The authors also propose a hybrid approach that consistently outperforms individual models and offer practical guidelines for selecting the right technique.

What Problem Does This Paper Address?

As Integrated Circuit (IC) designs become increasingly complex, accurately estimating critical metrics like power consumption, area, and delay is crucial for cost reduction and performance optimization. However, the sheer number of parameters involved in modern IC design – spanning technology, architecture, and physical layout – makes it incredibly challenging to develop precise analytical or physical models. This "curse of dimensionality" means that even small changes in many parameters can lead to complex, non-intuitive interactions that are hard to predict. Metamodeling, or creating surrogate models from actual design data, offers a promising solution. Yet, its application in IC design is still nascent, and designers lack a clear understanding of which metamodeling techniques, sampling strategies, and model complexities are best suited for different design estimation problems. This paper aims to fill that gap by systematically evaluating these factors.

Methods

The authors explore three popular metamodeling techniques:

  • Multivariate Adaptive Regression Splines (MARS): A tree-based method that uses piecewise linear functions.
  • Radial Basis Functions (RBF): A Gaussian process-based model that uses kernel functions.
  • Kriging (KG): Another Gaussian process-based model that incorporates spatial correlation.

They also introduce Hybrid Surrogate Modeling (HSM), a novel approach that combines the predictions of MARS, RBF, and KG by assigning weights to each, aiming to leverage the strengths of each individual model.

Two sampling strategies are compared:

  • Latin Hypercube Sampling (LHS): A standard technique for uniform sampling across parameter ranges.
  • Adaptive Sampling (AS): A more advanced method that intelligently selects new samples based on where the model's prediction error is highest or where the parameter space is least explored.

The study evaluates these methods across three distinct IC design estimation problems:

  1. Network-on-Chip (NoC) routers: Estimating area and power (low-dimensional).
  2. Standard cell delay and slew: Estimating under power delivery network (PDN) noise (high-dimensional).
  3. Clock Tree Synthesis (CTS): Estimating wirelength and buffer area (high-dimensional).

Datasets

The paper doesn't introduce new benchmark datasets but rather uses real-world IC design flows and tools to generate "golden" data for their three case studies:

  • NoC: Generated using Synopsys Design Compiler and Cadence SOC Encounter with TSMC65PLUS and TSMC45GS cell libraries, resulting in 1024 data points.
  • PDN: Generated using Synopsys HSPICE with TSMC65GPLUS design kit, creating three million data points.
  • CTS: Generated using Cadence SOC Encounter with TSMC65PLUS and TSMC45GS cell libraries, resulting in 1024 data points.

For training, they use subsets of these generated data points, with varying sample sizes (N), and a separate set for testing.

Key Insights and Ideas

  • Adaptive Sampling is King: AS consistently outperforms LHS, significantly reducing the number of samples needed for a given accuracy or achieving much lower errors for the same number of samples. This is because AS intelligently focuses on the most informative regions of the design space, unlike LHS's uniform approach.
  • Dimensionality Dictates Model Choice: The "best" metamodeling technique is not universal. For low-dimensional problems (D ≤ 5), Gaussian process models (RBF, KG) tend to be more accurate. However, as dimensionality increases (D > 7), tree-based models like MARS become significantly more effective, likely due to their better handling of multicollinearity.
  • Overfitting is a Real Danger: Simply adding more training samples doesn't always improve accuracy. Beyond a certain point, adding samples can lead to overfitting, where the model learns the noise in the training data rather than the underlying trend, thus increasing estimation errors on new data.
  • Hybridization Wins: The proposed Hybrid Surrogate Modeling (HSM) approach, which combines predictions from multiple models, consistently provides better accuracy than any single individual model, especially in challenging high-dimensional scenarios.

Key Results

  • Sampling Efficiency: Adaptive Sampling (AS) can reduce sample generation overhead by up to 64% or achieve up to 77% reduction in worst-case estimation error compared to LHS for the same accuracy.
  • Low-D Accuracy: For low-dimensional problems, RBF and KG can be up to 1.5x more accurate than MARS.
  • High-D Accuracy: For high-dimensional problems, MARS can be up to 6x more accurate than RBF and KG.
  • Hybrid Model Performance: HSM can improve estimation accuracy by up to 3x compared to the best individual surrogate model.
  • Multicollinearity Impact: High multicollinearity (where parameters are highly correlated) significantly degrades the performance of RBF and KG, while MARS remains relatively robust.

Limitations and Future Work

  • Generalizability of Guidelines: While the proposed guidelines are validated on the studied problems, the authors acknowledge that further studies are needed to confirm their generalizability across a wider range of IC design problems with varying dimensions.
  • Handling Ill-Conditioned Samples: The authors plan to explore methods for transforming ill-conditioned samples into well-conditioned ones to further improve the accuracy of all metamodeling techniques.
  • Exploring New ML Techniques: Future work will involve applying newer machine learning techniques, such as partial least-squares regression, for dimensionality reduction in IC design estimation problems.

Summary updated

BibTeX
@inproceedings{kahng2013enhanced,
  author = {Andrew B. Kahng and Bill Lin and Siddhartha Nath},
  title = {{Enhanced metamodeling techniques for high-dimensional IC design estimation problems}},
  booktitle = {DATE},
  year = {2013},
  doi = {10.7873/date.2013.371},
}
Figure