主讲人:苏仰锋 复旦大学教授
时间:2021年11月5日16:00
地点:三号楼332室
举办单位:数理学院
主讲人介绍:苏仰锋,博士,教授,博士生导师。曾任复旦大学数学科学学院信息与计算科学系系主任,中国计算数学学会常务理事,上海高校计算数学E-研究院特聘研究员,上海高校科学计算重点实验室客座成员,现任民进复旦大学邯郸总支委员会副主委。1982年考入复旦大学数学系计算数学专业,1986本科毕业,1989硕士毕业,1992.07博士毕业,留校任教。1996.5月至1998.4赴巴西里约热内卢联邦大学电工系访问,1999年被评为副教授,2005年被评为教授。主要研究数值代数的理论、方法和算法。特别强调在微电子领域中的应用。2012年研究成果“二阶Krylov子空间理论与集成电路分析中的模型降阶方法”获上海市自然科学一等奖。在SIAM J. Sci. Comput., SIAM J. Matrix Anal. Appl., IEEE Trans. Signal Process.,BIT等著名杂志发表30篇学术论文。与吴宗敏教授合著的《数值逼近》由科学出版社出版。
内容介绍:In modern Flat Panel Display (FPD) simulation, linear systems with unknowns are required to be solved. In the coming future, as the resolution of the FPD increases, the size of the linear system will become 109 or more. For a case with 4K resolution (3840 2160 pixels), explicit storage of the coefficient matrix needs around 500 GB of memory and the existing solver requires much more, which greatly limits the simulator application. We carefully exploit the structure of the linear systems in FPD simulation and propose an implicit storage format for the coefficient matrix. Based on the storage format, we construct an aggregation-based preconditioner and accelerate the matrix-vector multiplication in PCG. For low resolution cases, compared with the existing solvers, our algorithm requires only 10% of the memory while taking the comparable computational time. For high resolution cases, the existing solvers are out of memory, while our algorithm successfully solves the problem.