Research Of Multi-PVT SRAM Circuit Yield Analysis

Static Radom Acess Memory(SRAM)is greatly affected by the deviation of process parameters.Designers need to ensure that the circuit meets certain yield requirements.For the estimation of very low failure probability,the traditional Monte Carlo(MC)method needs tens of millions of circuit simulations to obtain accurate yield results.In addition,process voltage and temperature(PVT)deviation has a severe impact on the circuit.With the development of technology,the number of PVT combinations increases sharply.Most of the previous methods estimate the yield of fixed PVT,but the yield analysis under multiple PVT will consume a lot of simulation.A fast yield analysis method is designed.Firstly,under single PVT for SRAM yield under multiPVT,a yield analysis method based on importance sampling is designed.The model parameters are accurately estimated by penalized maximum likelihood estimation and expectation maximization algorithm,and the parameter initialization is completed by clustering algorithm based on cosine distance,which greatly improves the speed of yield estimation;Then,under multi PVT,a sample failure discrimination algorithm based on distribution adaptation is designed.Through distribution adaptation,samples under different PVT are mapped to the feature space with the smallest distribution distance,and the correlation between samples is used to avoid a large number of simulations and accelerate the yield analysis of multi PVT.The experimental scenarios include 18 dimensional SRAM bit-cell and 597 dimensional SRAM Column.Taking MC method as the gold standard,the hypersphere clustering sampling method,adaptive importance sampling method and multiple failure region importance sampling method are compared under single PVT,and the correlated Bayesian inference method and the independent evaluation method are compared under multi PVT.Under single PVT,the proposed method is 4167 times faster and 2092 times faster than MC method in two experimental scenarios,and 2.2 times ～8.7times faster than other benchmarking methods.In SRAM Column,the hypersphere clustering sampling method and adaptive importance sampling method cannot converge to the correct value.Under multi PVT,25 groups and 50 groups of PVT conditions were randomly generated in the experiment.In SRAM bit-cell experiments,the proposed method can accelerate more than 25000 times compared with MC method,and 4.2 times and 1.1 times faster than independent evaluation and correlated Bayesian inference respectively;In SRAM array experiments,the proposed method is more than 12800 times faster than MC method,4.7 times faster than independent evaluation and 1.35 times faster than correlated Bayesian inference method,respectively.