【作者】 周晨；

【导师】 谢爽；

【作者基本信息】 山东大学 ， 电子信息（专业学位）， 2024， 硕士

【摘要】 标准单元版图是集成电路设计的重要组成部分,版图的质量对于集成电路芯片的物理设计影响重大。随着集成电路产业的发展,标准单元的版图设计面临着巨大的挑战,目前标准单元的版图设计大都是手工进行,耗费了大量时间和人力。针对这一问题,论文提出了标准单元版图布局的自动化算法,主要围绕版图的设计优化展开研究:(1)论文从标准单元的设计流程和版图设计的原理与规则出发,提出了单元版图优化的方式,从面积、线长、可布线性三方面进行多目标优化,确定自动化实现布局的目标函数。(2)论文讨论了模拟退火算法的原理以及算法在标准单元版图设计中的应用,从电路结构的角度出发,提出了基于CCB的标准单元分块算法,将分块算法与模拟退火算法结合,充分发挥了模拟退火在运行速度方面的优势。(3)论文利用强化学习追求最大回报的特点,对标准单元进行自动化布局,论文对比了各类强化学习算法的优劣,提出使用策略梯度算法对标准单元进行自动化布局,建立了强化学习模型,并使用蒙特卡洛方法对神经网络进行了训练。论文使用28nm的标准单元库进行了试验,标准单元库包括基本组合逻辑单元,如AND、OR等;时序逻辑单元,如D触发器,锁存器等;物理填充单元,如Buffer等。分别使用两种算法进行标准单元自动化布局,实验结果表明,两种算法在小单元中均取得了较好的表现,其中基于分块算法的标准单元自动化布局较强化学习的算法速度更快,且在大单元中的布局仍然取得了较好的结果。更多还原

【Abstract】 Standard cell layout is an important part of IC design,and the quality of layout has a significant impact on the physical design of IC chips.With the development of IC industry,the layout design of standard cell is facing great challenges.At present,the layout design of standard cell is mostly carried out manually,which consumes a lot of time and labor.Aiming at this problem,the thesis proposes an automated algorithm for standard cell layout,mainly focusing on the design optimization of the layout:(1)Starting from the design process of standard cells and the principles and rules of layout design,the thesis proposes a way to optimize the cell layout,carry out multi-objective optimization from the aspects of area,line length,and routability,and determine the objective function of the automation to realize the layout.(2)The thesis discusses the principle of simulated annealing algorithm and the application of the algorithm in the layout design of standard cell,and from the perspective of circuit structure,it proposes the standard cell blocking algorithm based on CCB,and combines the chunking algorithm with the simulated annealing algorithm to give full play to the advantages of simulated annealing in terms of operation speed.(3)The thesis utilizes the characteristics of reinforcement learning in pursuit of maximum return to automate the layout of standard cell.The thesis compares the advantages and disadvantages of various types of reinforcement learning algorithms,proposes the use of policy gradient algorithm to automate the layout of standard cell,establishes a reinforcement learning model,and uses Monte Carlo methods to train the neural network.The thesis conducted experiments using a 28 nm standard cell library,which includes basic combinational logic cell,such as AND,OR,etc.;timing logic cell,such as D flip-flop,latch,etc.;and physical filler cell,such as Buffer.Two algorithms are used to automate the layout of standard cell,and the experimental results show that both algorithms achieve better performance in small cells,in which the automated layout of standard cell based on the blocking algorithm is faster than the algorithm of reinforcement learning,and the layout of large cells still achieves better results.更多还原