论文部分内容阅读
随着芯片制造工艺的进步,工艺参数、供电电压及片上温度(Process,voltage,and temperature,PVT)等设计参数扰动已成为芯片设计过程的棘手问题,其所产生的性能指标间相关性将导致芯片参数成品率显著下降。但是,当前芯片参数成品率预测算法主要局限于单一性能指标成品率预测或对多个单性能指标成品率进行均衡优化,而不能同时针对多个性能指标约束进行多元参数成品率预测,易造成参数成品率精度缺失。基于以上问题,本文将多个性能指标同时作为约束条件,提出一种芯片多元参数成品率预测方法。该方法首先考虑PVT参数扰动,利用自适应弹性网(Adaptive elastic net,AEN)对芯片性能指标进行建模。然后,基于乘法定理及马尔科夫链蒙特卡罗法,通过求解累积分布函数(Cumulative distribution function,CDF)对单一性能指标的芯片参数成品率进行预测。最后,同时考虑多个芯片性能指标约束,根据Copula方法准确预测芯片多元参数成品率。实验结果表明,本文方法可以在指定性能指标约束下对芯片多元参数成品率进行有效预测,并可为芯片设计人员提供任意性能指标约束下的多元参数成品率预测曲面。
With the advancement of the chip manufacturing process, disturbance of design parameters such as process parameters, supply voltage and PVT has become a thorny issue in the chip design process, and the correlation among the performance indexes generated will lead to Chip parameters significantly reduced the yield. However, the current chip parameter yield prediction algorithm is mainly limited to the prediction of the yield of a single performance index or the optimization and optimization of the yield of multiple single-performance indexes, and can not predict the yield of multiple parameters at the same time with multiple performance constraints. Yield accuracy is missing. Based on the above problems, this paper will be a number of performance indicators at the same time as a constraint condition, a chip multivariate parameter yield prediction method. The method first considers the PVT parameter perturbation and models the chip performance using an adaptive elastic net (AEN). Then, based on the multiplicative theorem and Markov chain Monte Carlo method, the chip parameters yield of a single performance index is predicted by solving the Cumulative distribution function (CDF). Finally, considering multiple chip performance constraints, the chip multivariate parametric yield is accurately predicted according to the Copula method. Experimental results show that this method can effectively predict chip multivariate parameter yield under the constraints of specified performance index and provide chip designers with multivariable parameter yield prediction surface under the constraint of arbitrary performance index.