Seeking the Yield Barrier: High-Dimensional SRAM Evaluation Through Optimal Manifold
OPTIMIS method improves efficiency by 3.5x and accuracy by 3x in high-dimensional SRAM evaluation.
Key Findings
Methodology
This paper introduces a novel yield estimation method, OPTIMIS, which combines the advantages of importance sampling and surrogate models using neural coupling flow as the proposal distribution. The method derives the optimal manifold and suboptimal manifold (optimal hypersphere), proposing onion sampling to efficiently sample from the failure distribution. OPTIMIS demonstrates excellent performance in high-dimensional SRAM evaluation with robustness and consistency.
Key Results
- In a 108-dimensional SRAM circuit, OPTIMIS is 131.89x faster than MC, with accuracy over 10x better.
- In a 569-dimensional circuit, OPTIMIS achieves a 273.82x speedup with only 0.25% error.
- In a 1093-dimensional circuit, OPTIMIS shows a 185.78x speedup with just 2.71% error.
Significance
This study provides an efficient and robust method for yield estimation in high-dimensional SRAM circuits, addressing the computationally intensive issue of traditional Monte Carlo methods. By integrating importance sampling and machine learning techniques, OPTIMIS excels in handling high-dimensional data, offering broad industrial application potential.
Technical Contribution
OPTIMIS significantly enhances yield estimation efficiency and accuracy by introducing the optimal manifold and neural coupling flow. Compared to existing methods, OPTIMIS offers significant advantages in sampling efficiency and model updates, better fitting the failure distribution in high-dimensional spaces.
Novelty
OPTIMIS is the first method to apply the optimal manifold concept to SRAM yield estimation, providing new theoretical guarantees and engineering possibilities through the combination of onion sampling and neural coupling flow.
Limitations
- In some high-dimensional cases, onion sampling may lead to initial estimation bias.
- Neural coupling flow requires a large amount of data, potentially increasing computational costs.
Future Work
Future research could explore applying OPTIMIS to larger-scale circuits and optimizing the training process of neural coupling flow to reduce computational resource requirements.
AI Executive Summary
With the advancement of integrated circuit technology, the shrinking size of microelectronic devices makes random process variations a critical factor in circuit design. Traditional Monte Carlo methods are computationally intensive and inadequate for modern circuit design needs.
This paper proposes a novel yield estimation method, OPTIMIS, which combines the advantages of importance sampling and surrogate models using neural coupling flow as the proposal distribution. The method derives the optimal manifold and suboptimal manifold (optimal hypersphere), proposing onion sampling to efficiently sample from the failure distribution. OPTIMIS demonstrates excellent performance in high-dimensional SRAM evaluation with robustness and consistency.
Experimental results show that OPTIMIS significantly outperforms existing methods in handling high-dimensional SRAM circuits, offering broad industrial application potential. Future research could explore applying OPTIMIS to larger-scale circuits and optimizing the training process of neural coupling flow to reduce computational resource requirements.
Deep Analysis
Background
With the advancement of integrated circuit technology, the shrinking size of microelectronic devices makes random process variations a critical factor in circuit design. Traditional Monte Carlo methods are computationally intensive and inadequate for modern circuit design needs. Recently, importance sampling and machine learning techniques have been introduced to improve yield estimation efficiency.
Core Problem
Accurately estimating failure probability in high-dimensional SRAM circuits is challenging. Traditional methods are computationally intensive and struggle with low-yield problems. A more efficient and robust method is needed to address this issue.
Innovation
OPTIMIS combines the advantages of importance sampling and surrogate models, significantly improving yield estimation efficiency and accuracy by introducing the optimal manifold and neural coupling flow. Onion sampling efficiently samples from the failure distribution.
Methodology
- �� Derive the optimal manifold and suboptimal manifold (optimal hypersphere).
- �� Propose onion sampling to efficiently sample from the failure distribution.
- �� Use neural coupling flow as the proposal distribution, learning failure probability from samples.
Experiments
Experiments were conducted on 108, 569, and 1093-dimensional SRAM circuits using Synopsys Design Compiler and Cadence Virtuoso design tools. The target failure rate was set at 10^-5, with MC methods as the baseline.
Results
In the 108-dimensional circuit, OPTIMIS is 131.89x faster than MC, with only 0.21% error. In the 569-dimensional circuit, it achieves a 273.82x speedup with just 0.25% error. In the 1093-dimensional circuit, the speedup is 185.78x with only 2.71% error.
Applications
OPTIMIS has broad application potential in yield estimation for high-dimensional SRAM circuits, significantly improving efficiency and accuracy while reducing computational resource requirements.
Limitations & Outlook
Onion sampling may lead to initial estimation bias, and neural coupling flow requires a large amount of data, potentially increasing computational costs. Future research could optimize the training process.
Plain Language Accessible to non-experts
Imagine you're in a kitchen trying to quickly find all the possible factors that could spoil your food. OPTIMIS acts like a smart assistant that helps you quickly identify these factors and provides solutions. By combining multiple methods, it works efficiently in complex environments to ensure food safety and taste.
ELI14 Explained like you're 14
Imagine you're playing a game where you need to find all the hidden treasures on a map. OPTIMIS is like a super helper that quickly finds these treasures and gives you tips on how to find them faster. It combines multiple methods to help you perform better in the game!
Glossary
SRAM (Static Random-Access Memory)
A type of memory that uses bistable latching circuitry to store data, commonly used in cache.
Used in this paper to evaluate high-dimensional circuit yield.
Importance Sampling
A statistical technique for estimating expectations by sampling from a proposal distribution.
Used to improve yield estimation efficiency.
Neural Coupling Flow
A deep learning model used to generate complex distributions through invertible transformations.
Used in OPTIMIS to learn failure probability distributions.
Onion Sampling
A method for efficiently sampling from the failure distribution by peeling layers of a hypersphere.
Used to improve sampling efficiency and accuracy.
Optimal Manifold
A theoretical concept for guiding proposal distribution design, combining the advantages of importance sampling and surrogate models.
Used in OPTIMIS to enhance yield estimation performance.
Open Questions Unanswered questions from this research
- 1 How can OPTIMIS be applied to larger-scale circuits?
- 2 How can the training process of neural coupling flow be optimized to reduce computational resource requirements?
- 3 How does onion sampling perform in more complex failure distributions?
Applications
Immediate Applications
High-dimensional SRAM Circuit Evaluation
OPTIMIS can be used for rapid yield estimation of high-dimensional SRAM circuits, reducing computational resource requirements.
Long-term Vision
Complex Circuit Design Optimization
OPTIMIS can be used to optimize complex circuit designs, improving efficiency and accuracy, and driving industry development.
Abstract
Being able to efficiently obtain an accurate estimate of the failure probability of SRAM components has become a central issue as model circuits shrink their scale to submicrometer with advanced technology nodes. In this work, we revisit the classic norm minimization method. We then generalize it with infinite components and derive the novel optimal manifold concept, which bridges the surrogate-based and importance sampling (IS) yield estimation methods. We then derive a sub-optimal manifold, optimal hypersphere, which leads to an efficient sampling method being aware of the failure boundary called onion sampling. Finally, we use a neural coupling flow (which learns from samples like a surrogate model) as the IS proposal distribution. These combinations give rise to a novel yield estimation method, named Optimal Manifold Important Sampling (OPTIMIS), which keeps the advantages of the surrogate and IS methods to deliver state-of-the-art performance with robustness and consistency, with up to 3.5x in efficiency and 3x in accuracy over the best of SOTA methods in High-dimensional SRAM evaluation.