iVAMS 3.0: Hierarchical-Machine-Learning-Metamodel-Integrated Intelligent Verilog-AMS for Ultra-Fast, Accurate Mixed-Signal Design Optimization

TL;DR

This paper introduces a hierarchical machine learning metamodel integrated with an optimization flow for ultra-fast, accurate analog/mixed-signal circuit design, significantly reducing optimization time and improving variability management.

Contribution

It presents a novel Kriging-bootstrapped ANN metamodel combined with PSO for rapid, scalable, and accurate AMS circuit optimization, addressing accuracy and speed trade-offs.

Findings

Kriging-bootstrapped ANN is 24X faster than simple ANN.

The approach effectively reduces PLL variability.

Significant reductions in mean and standard deviation of PLL characteristics.

Abstract

Analog/Mixed-Signal (AMS) circuits and systems continually present significant challenges to designers with the increase of design complexity and aggressive technology scaling. This is due to the large number of design factors and parameters that must be taken into account as well as the process variations which are prominent in nano-CMOS circuits. Design optimization techniques while presenting an accurate and fast design flow which can perform design optimization in reasonable time are still lacking. Even with techniques such as metamodeling that aid the design phase, there is still the need to improve them for accuracy and time cost. As a trade-off of the accuracy and speed, this paper presents a design flow for ultra-fast variability-aware optimization of nano-CMOS based physical design of analog circuits. It combines a Kriging bootstrapped Artificial Neural Network (ANN) metamodel with a Particle Swarm Optimization (PSO) based algorithm in the design optimization flow. The Kriging bootstrapped ANN metamodel provides a trade-off between analog-quality accuracy and scalability and can be effectively used for large and complex AMS circuits. The proposed technique uses Kriging to bootstrap target samples used for the ANN training. This introduces Kriging characteristics, which account for correlation effects between design parameters, to the ANN. The effectiveness of the design flow is demonstrated using a PLL as a case study with as many as 21 design parameters. It is observed that the bootstrapped Kriging metamodeling is 24X faster than simple ANN metamodeling. The layout optimization for such a complex circuit can be performed effectively in a short time using this approach. The optimization flow could achieve significant reductions in the mean and standard deviation of the PLL characteristics. Thus, the proposed research is a major contribution to design for cost.