A Research On An AdaBoost-Based Algorithm For The Failure Rate Of SRAM Cell

As deep submicron technology advances, improvement in integrated circuit (IC) performance and cost has been achieved largely by transistor scaling. The continuous scaling down of critical dimension and the highly integration of IC pushes the design methodology of IC to update itself. The design methodology of IC has experienced the revolution of device based generation to interconnect based generation, and now it’s entering the third generation of Design for Manufacturing (DFM) and Design for Yield (DFM). As the feature size going down, circuit reliability under process variations is an area of growing concern. The ability to efficiently obtain an accurate estimation of yield is becoming a central issue.Currently, Static Random Access Memory (SRAM) has been the main component in a large amount of system on chip. To assure the device count per chip, SRAM cells are generally designed with minimum-size device and can be significantly impacted by process variations. Because of the highly replicated circuits of SRAM, the failure rate of a single cell should be quite low to assure the yield rate, which is call rare statistical event. The key point of predicating the yield rate is performing a precise simulation of the failure rate.It’s tougher for the failure rate of a single cell as the exponential growth of SRAM’s storage, traditional sampling methods perform poorly and are no longer available. The dissertation focuses on the estimation of the rare probability of SRAM cell caused by process variations. It builds models for the electrical parameters influenced by process variations and combines the machine learning and extreme value theory. An algorithm of Adaptive Boosting (AdaBoost) Based Quasi Importance Sampling is proposed in this thesis.There are two kinds of methods for the estimation of the failure rate of SRAM cell:deterministic methods and statistical analysis. The dissertation focuses on the later one and gives a detail introduction and analysis for it. The new algorithm proposed in this thesis starts from these two followings:AdaBoost classifier is used for filtering the samples to avoid SPICE simulation for the useless samples; Improve the traditional IS method by the method of minimizing the cross entropy, which increase the convergence rate of the algorithm. Compared to the existing state-of-the-art techniques, this algorithm runs faster for the same accuracy.