Statistical Performance Modeling of SRAMs

Yield analysis is a critical step in memory designs considering a variety of performance constraints. Traditional circuit level Monte-Carlo simulations for yield estimation of Static Random Access Memory (SRAM) cell is quite time consuming due to their characteristic of low failure rate, while statistical method of yield sensitivity analysis is meaningful for its high efficiency. This thesis proposes a novel statistical model to conduct yield sensitivity prediction on SRAM cells at the simulation level, which excels regular circuit simulations in a significant runtime speedup. Based on the theory of Kriging method that is widely used in geostatistics, we develop a series of statistical model building and updating strategies to obtain satisfactory accuracy and efficiency in SRAM yield sensitivity analysis. Generally, this model applies to the yield and sensitivity evaluation with varying design parameters, under the constraints of most SRAM performance metric. Moreover, it is potentially suitable for any designated distribution of the process variation regardless of the sampling method.