Research On Thermal Analysis Method Of Chip Based On CUDA

Before the microelectronic chip is manufactured,it is important to accurately estimate the thermal distribution and identify the local hotspots within the chip to verify and predict the performance and reliability of the chip.Because VLSI chips work at extremely fast clock speeds,they can cause catastrophic thermal breakage.In the accurate estimation of chip thermal analysis,there have been many papers that the heat spreader and heat sink have the same dimensions as the chip.However,if the geometry of the heat spreader and heat sink is not taken into consideration,the experimental results show that it will cause tens of degrees of temperature error.In this thesis,the serial random walk algorithm based on pyramid-shape thermal analysis model is studied and improved.First of all,the pre-computed probabilities for the scalable transition domain is mapped to the pyramid-shape model,and then the subdomain coding and masking techniques are used to simplify the jump process.For a thermal model with a high degree of dispersion,the computational efficiency of the serial random walk algorithm is very low because the number of random jumps will increase dramatically as the degree of discretization increases.Firstly,this thesis establishes the parallelization model of random walking algorithm based on CUDA.Secondly,the parallel random walk algorithm is studied deeply,which solves the problem of initialization of three-dimensional array in GPU,random number processing problem,locking mechanism and warp divergence problem.Thirdly,the parallel algorithm is optimized and the optimal configuration is obtained.And the comparison between the serial algorithm and parallel algorithm in the implementation of efficiency and accuracy is carried out.Finally,the time complexity of the serial and parallel random walk algorithm is analyzed.The experimental results show that compared with the serial random walk algorithm,the parallel random walk algorithm designed in this thesis can improve the execution speed under the premise of keeping the accuracy.The chip Structure1 acceleration ratio up to 9.97 times,the Structure2 acceleration ratio up to 9.13 times.The higher the degree of dispersion of the model,the better the acceleration effect of the algorithm.