Research And Implementation On A Hybrid Random Walk Algorithm For 3-D Thermal Analysis Of Integrated Circuits

Thermal analysis plays a very important role in high-performance very large scale integration(VLSI)design.Fast and accurate thermal analysis of integrated circuits(ICs)is crucial for both the sign-off stage and design-time circuit optimization.It is performed during the sign-off stage for performance and reliability verification.In addition,it is also indispensable for many design-time circuit optimizations.The finite volume discretization has been employed to thermal analysis of IC,which transforms the problem into a linear equation system.With the volume discretization,the thermal profile of the whole simulated domain can be obtained by solving the linear equation system in most prior works.However,the chip-level thermal analysis that solves for the thermal profile of an entire chip is not always required,because the temperatures of some target hot-spots in IC are often desired.For accurate calculations of the on-chip thermal profile,it's necessary to simulate a chip together with its heat sink components in chip package.In this thesis,both a hybrid random walk method and its corresponding model are presented for the realistic pyramid-shaped IC thermal model including heat sink and heat spreader,where the hybrid random walk method combines the advantages of the generic random walk method and the floating random walk method.Not only does the hybrid random walk method have the desirable property of simulating localized hot-spots,but also it largely reduces its runtime for 3-D thermal analysis of the non-rectangular thermal model.What's more,the hybrid random walk method based on the techniques handling the Neumann boundary and Convective boundary in the pyramid thermal model is also presented to further improve the efficiency.Finally,the parallel computing technique is applied to greatly accelerate the computation speed of the hybrid random walk method on the multicore/multi-CPU platform.To validate the efficiency and accuracy of the proposed techniques,numerical experiments are carried out with several IC test cases.They demonstrate that the hybrid random walk method achieves up to 15.9X speedup over the generic random walk method.And the hybrid random walk method based on the techniques handling the Neumann boundary brings 30.1X speedup over the generic random walk method.Furthermore,the hybrid random walk method based on the techniques handling the Neumann boundary and Convective boundary brings 44.4X speedup over the generic random walk method.The experiments on a 6-core CPU machine show that the parallel hybrid random walk method is about 5X faster than its serial-computing version.