基于自适应增强的静态随机存储单元失效率快速仿真算法  

An AdaBoost-Based Fast Algorithm for SRAM Cell Failure Rate Analysis

作　　者：韦峰[1] 朱恒亮[1] 曾璇[1] 

机构地区：[1]复旦大学专用集成电路与系统国家重点实验室,上海201203

出　　处：《复旦学报（自然科学版）》2014年第3期350-358,364,共10页Journal of Fudan University：Natural Science

基　　金：国家重点基础研究发展计划(2011CB309701);国家自然科学基金(61106032;91330201;61125401);国家十二五科技重大专项项目(2014ZX02301001-005-002);上海市领军人才资助项目

摘　　要：随着纳米工艺的不断推进,由随机工艺参数偏差引起的电路稳定性问题日益突出.静态存储器由大规模高度重复性电路单元结构组成,要保证整个芯片工作正常,要求单个单元电路失效的概率极低,相关的失效事件属于极端概率事件.传统的蒙特卡洛采样在产生足够的有效采样点和精确捕捉实际失效区域分布方面存在诸多缺陷.本文采用的自适应增强(AdaBoost)方法是一种新的统计学习技术,通过样本学习构建一个强分类器,可以过滤掉大量无效的采样点,使采样点更集中的分布在失效区域中.另外,本文引入交叉熵方法改进传统的重要性采样算法,达到更快的收敛速度.该方法综合了数据挖掘和极值理论的思想,能够快速精确地捕捉失效区域实际分布,相比于标准的蒙特卡洛采样方法,在不损失精度的情况下效率提高了1000倍以上.As deep submicron technology advances, circuit reliability under statistical process variation is becoming increasingly critical. Static Random Access Memory is generally designed with minimum-size devices with highly replicated circuit structure whose failure rate is an area of growing concern. Traditional Monte Carlo sampling method suffers from generating efficient samples and estimating the practical failure region＇s distribution function. The AdaBoost learning method is a new statistical learning technique, constructing a strong classifier with training samples. This technique could filter most of the trivial samples and make the samples more centered in failure region. In this paper, we propose an AdaBoost-based improved importance sampling method to accurately estimate the distribution function of the failure region. This algorithm combines the data mining and extra-value theory. Experimental results shows it achieves 1 000 times speedup over the MC method with little precision loss.

关 键 词：静态随机存储器 深亚微米 蒙特卡洛采样 随机工艺偏差 重要性采样 ADABOOST方法 

分 类 号：TM344.1[电气工程—电机]