Multidisciplinary heat generating cell placement optimization using genetic algorithm and artificial neural networks

A multidisciplinary optimization methodology for placement of heat generating cells on integrated circuit chips is presented. The methodology includes thermal and wiring length criteria, which are optimized simultaneously using the genetic algorithm. An effective thermal performance prediction methodology, which is based on a combination of a superposition method and artificial neural networks (ANNs) is developed for this study. Radial bases function network is trained using finite element models and are integrated to predict the temperature distribution on a silicon chip due to a single heat generating transistor in a cell. A superposition method is used to calculate the temperature distribution on a silicon chip due to multiple heat generating transistors. Using the artificial neural networks, the predicted temperature distribution in the silicon chip is obtained in a much shorter time than with a full finite element model and with comparable accuracy. The main advantage of the present multi-disciplinary design and optimization methodology is its ability to handle multiple design objectives simultaneously for optimized placement of heat generating cells. This unique capability distinguishes the present methodology from existing ones. To demonstrate its capabilities, the present methodology is applied to a test case involving placement optimization of multiple heat generating cells on a silicon chip. The results indicate that a substantial reduction in maximum junction temperature combined with optimized wiring length can be obtained. The placement optimization results indicate that block placement rather than cell placement improves the maximum temperature in a silicon chip.