Dealing with Aging and Yield in Scaled Technologies

Wei Ye, Mohamed Baker Alawieh, Che-Lun Hsu, Yibo Lin, David Z. PanEmbedded 2020 · DOI · OpenAlex

Abstract

This chapter reviews recent practices of tackling aging and yield issues in VLSI design related to shrinking technology processes. Different fundamental effects such as device aging, interconnect electromigration, and process variations are investigated with the state-of-the-art techniques for modeling and optimization. The presented techniques vary from analytical approaches to machine learning, and often require cross-layer information feedback for robust design cycles.

Introduction: The Reliability Crisis in Sub-45nm Nodes

As semiconductor manufacturing pushes into the sub-45nm realm, traditional "fix-it-later" design approaches are becoming obsolete. In the era of aggressive technology scaling, reliability is no longer a constant. Two major hurdles threaten the viability of modern chips: aging and yield loss.

Aging effects, such as Bias Temperature Instability (BTI), Hot Carrier Injection (HCI), and Electromigration (EM), cause circuits to degrade or fail over their operational lifetime. Historically, these failures followed a predictable "bathtub curve," but in advanced nodes, these effects have become increasingly stochastic.

The bathtub curve illustrates the three phases of device failure: infant mortality, normal lifetime, and wear-out.

Furthermore, process variations lead to performance mismatches or "lithography hotspots"—patterns that fail to print correctly using current 193nm lithography. Because late-stage fixes like Engineering Change Orders (ECOs) are prohibitively expensive, modern VLSI design is shifting these concerns "left" into the placement and synthesis stages.

Methods: From Physics to Machine Learning

The paper surveys a spectrum of techniques designed to combat these challenges, moving from physics-based analytical models to cutting-edge machine learning (ML):

  • Reliability Modeling: The review discusses Reaction-Diffusion (R-D) and Charge Trapping (CT) models for BTI, alongside Black’s Equation for Electromigration.
  • Machine Learning for EM: A two-stage logistic regression approach is highlighted for predicting signal electromigration hotspots at the placement stage, long before routing is finalized.
  • Sparse & Semi-Supervised Modeling: To model Analog Mixed-Signal (AMS) performance efficiently, the paper explores Bayesian "spike and slab" feature selection and co-learning frameworks that leverage small labeled datasets alongside large unlabeled ones.
  • Bayesian Optimization: Gaussian Processes are used as surrogate models for "black-box" performance optimization, finding optimal parameters with minimal simulation cycles.
  • Lithography Hotspot Detection: The transition from simple pattern matching to Deep Neural Networks (DNNs) and Litho-GPA—a framework using Gaussian Process Regression to provide confidence guarantees—is a key focus.

Key Insights: Embracing Stochasticity and Cross-Layer Synergy

A central theme of recent research is the transition from deterministic to stochastic modeling. In sub-45nm nodes, identical transistors may age differently due to the random nature of charge trapping and interface state generation.

Comparison of threshold voltage shifts (ΔVth) over time. (a) Large-scale averaging (N=800) masks individual variations, while (b) small-scale devices (N=12) exhibit highly stochastic aging behavior.

Other critical insights include:

  • Self-Healing Interconnects: Modern AC-EM models now account for the "self-healing" effect in signal wires, where material backflow during alternating current cycles partially repairs electromigration damage.
  • Trustworthy AI in CAD: The Litho-GPA framework exemplifies a move toward "trustworthy" ML, where uncertain predictions are flagged for traditional (but slow) lithography simulation, ensuring accuracy without sacrificing speed.
  • Cross-Layer Synergy: The authors argue that robustness cannot be solved at the physics layer alone; information must flow from the physical layer up to architecture and logic synthesis for truly robust design.

Key Results and Future Outlook

The application of these advanced modeling techniques has yielded significant improvements in design efficiency and accuracy:

  • Efficiency: Semi-supervised learning for AMS modeling achieved a 3.6× speedup over traditional sparse regression.
  • Data Efficiency: Active learning for hotspot detection reached state-of-the-art accuracy using only 17% of the typical training data.
  • Accuracy: The Litho-GPA framework reduced false alarms by 28% while maintaining high detection accuracy.

Despite these gains, challenges remain. The computational overhead of cross-layer feedback is significant, and the interpretability of DNNs in hotspot detection requires further work. Future research must continue to balance the conflicting trade-offs between aging mitigation, power consumption, and leakage in increasingly complex scaled technologies.

Summary updated

Abstract

Abstract This chapter reviews recent practices of tackling aging and yield issues in VLSI design related to shrinking technology processes. Different fundamental effects such as device aging, interconnect electromigration, and process variations are investigated with the state-of-the-art techniques for modeling and optimization. The presented techniques vary from analytical approaches to machine learning, and often require cross-layer information feedback for robust design cycles.
BibTeX
@inproceedings{ye2020dealing,
  author = {Wei Ye and Mohamed Baker Alawieh and Che-Lun Hsu and Yibo Lin and David Z. Pan},
  title = {{Dealing with Aging and Yield in Scaled Technologies}},
  booktitle = {Embedded},
  year = {2020},
  doi = {10.1007/978-3-030-52017-5\_17},
}
Figure