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

发布时间：2018-07-06 15:32

本文选题：静态随机存储器 + 深亚微米　；参考：《复旦学报(自然科学版)》2014年03期

【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.
[Abstract]:With the continuous advancement of nanotechnology, the problem of circuit stability caused by random process parameter deviation has become increasingly prominent. Static memory is composed of large and highly repetitive circuit unit structure. To ensure that the whole chip is working normally, the probability of failure of single unit circuit is very low, and the related failure events belong to extreme probability events. The traditional Monte Carlo sampling has many defects in producing enough effective sampling points and accurately capturing the distribution of actual failure regions. The adaptive enhancement (AdaBoost) method used in this paper is a new statistical learning technique. A strong classifier is constructed through sample learning, which can filter out a large number of invalid sampling points and make the sampling point. More concentrated in the failure area. In addition, this paper introduces the cross entropy method to improve the traditional importance sampling algorithm to achieve faster convergence speed. This method combines the idea of data mining and extremum theory, and can quickly and accurately capture the actual distribution of the failure region, compared to the standard Monte Carlo sampling method, it is not lost. In the case of precision, the efficiency is increased by more than 1000 times.
【作者单位】：复旦大学专用集成电路与系统国家重点实验室;
【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
【分类号】：TP333

【相似文献】