Efficient SRAM yield optimization with mixture surrogate modeling  

作　　者：蒋中建 叶佐昌 王燕 

机构地区：[1]Tsinghua National Laboratory for Information Science and Technology,Institute of Microelectronics,Tsinghua University

出　　处：《Journal of Semiconductors》2016年第12期64-69,共6页半导体学报（英文版）

摘　　要：Largely repeated cells such as SRAM cells usually require extremely low failure-rate to ensure a mod- erate chi yield. Though fast Monte Carlo methods such as importance sampling and its variants can be used for yield estimation, they are still very expensive if one needs to perform optimization based on such estimations. Typ- ically the process of yield calculation requires a lot of SPICE simulation. The circuit SPICE simulation analysis accounted for the largest proportion of time in the process yield calculation. In the paper, a new method is proposed to address this issue. The key idea is to establish an efficient mixture surrogate model. The surrogate model is based on the design variables and process variables. This model construction method is based on the SPICE simulation to get a certain amount of sample points, these points are trained for mixture surrogate model by the lasso algorithm. Experimental results show that the proposed model is able to calculate accurate yield successfully and it brings significant speed ups to the calculation of failure rate. Based on the model, we made a further accelerated algo- rithm to further enhance the speed of the yield calculation. It is suitable for high-dimensional process variables and multi-performance applications.Largely repeated cells such as SRAM cells usually require extremely low failure-rate to ensure a mod- erate chi yield. Though fast Monte Carlo methods such as importance sampling and its variants can be used for yield estimation, they are still very expensive if one needs to perform optimization based on such estimations. Typ- ically the process of yield calculation requires a lot of SPICE simulation. The circuit SPICE simulation analysis accounted for the largest proportion of time in the process yield calculation. In the paper, a new method is proposed to address this issue. The key idea is to establish an efficient mixture surrogate model. The surrogate model is based on the design variables and process variables. This model construction method is based on the SPICE simulation to get a certain amount of sample points, these points are trained for mixture surrogate model by the lasso algorithm. Experimental results show that the proposed model is able to calculate accurate yield successfully and it brings significant speed ups to the calculation of failure rate. Based on the model, we made a further accelerated algo- rithm to further enhance the speed of the yield calculation. It is suitable for high-dimensional process variables and multi-performance applications.

关 键 词：yield optimization process variations design variations mixture surrogate model statistical analysis importance sampling 

分 类 号：TP333[自动化与计算机技术—计算机系统结构]