SPEEDING‐UP OF CONVERGENCE OF GUMMEL ITERATIONS FOR TRANSIENT SIMULATION

Abstract
Enhanced efficiency of the proposed modification of the Gummel decoupled method enables significant reduction of memory requirements in comparison to the conventionally used Newton‐like methods with approximately equal CPU time expense. For 10,000 mesh nodes it only requires 4.4 Mbytes of virtual memory (compare to 62 Mbytes for PISCES‐2B). The method exhibits good stability and convergence rate and has significant advantages compared to the other often used half‐implicit scheme. Low memory requirements and efficiency make the method attractive for future 3‐D applications.