Gaussian Process Based Sampling Algorithm For Analog Circuits Yield Analysis

With the advanced process technology,the impact of process variations on chip yield is more severe and analog circuits yield analysis becomes one of research hotspots.Traditional yield analysis methods become inefficient for analog circuits scenario.Traditional Monte Carlo method requires intolerable analysis time,importance sampling method proposes fixed sampling distributions which are hard to adjust for different analog circuits and the approximation error of meta-model method cannot be mathematically evaluated.Thereby,it is necessary to develop fast Monte Carlo yield analysis method for analog circuits.In this thesis,Gaussian process meta-model importance sampling(GMIS)is designed to solve multi-specification analog circuits yield analysis problem through combing meta-model and importance sampling method.GMIS first builds Gaussian process meta-models to approximate mapping relationships between input and output space.Then a quasi-optimal sampling distribution is constructed based on the meta-model for locating failure region boundaries.Finally parallel coupled Markov Chain Monte Carlo is implemented to extract sampling from proposed distribution.GMIS also designs dynamic learning framework which updates the training set and improves model prediction accuracy.The accuracy and efficient of this framework are proved by formula derivation.Proposed GMIS is verified on charge pump,two-stage transconductance amplifier and four-stage voltage-controlled ring oscillator.For charge pump,GMIS estimation has 2.86%relative error and accelerates 133 times compared with Monte Carlo method.For two-stage transconductance amplifier with four design specifications,GMIS uses 5 times and 27 times less samples compared with several recent methods and reaches 6.7% relative error.For voltagecontrolled ring oscillator with 120 process variables,GMIS accelerates 125 times compared with Monte Carlo method with only 3.7% relative error.Combined with space exploration algorithm,GMIS method can be further extended to analog yield optimization algorithm,which is able to automatically find the best yield-aware design point for stable and robust analog circuits.