| The full name of GPU is Graphic Processing unit. Since the development of the electronic computers, GPU is always the basis configuration. At the elementary stage of the development of computer, the performance of GPU is weak, which only undertakes the role of displaying figures. As the development of computers, the performance of GPU make spurt of progresses. Moreover, resorting to its special structure, GPU makes the floating decimals of GPU and the parallel computing ability ten times or even a hundred times better than central processing unit (CPU). Because of the weak computational ability of the single CPU in the traditional scientific computing, it obviously slows the computational rate. Although multi-CPU parallel computing can improve the computing rate, large scale computer clusters are too resources consumption. GPU can just in time supply the shortcomings of traditional calculations. Thus, GPU will play a very impartment role in future scientific computations.In this paper, we adopt the acceleration strategy of GPU to compute the energies and of the grads of (Al2O3)n clusters, and we compute the acceleration ratio of GPU and traditional CPU. We apply the acceleration strategy of GPU to the global optimizations of (Al2O3)n clusters.The main contents are as follows:1、The calculations of energy and gradient based on GPUFor (Al2O3)n clusters, we regard them as rigid models. The potential energy function has four terms, and the calculations of them are executed in GPU. Finally, we obtain a high acceleration ration. Conclusions are as follows:there are three different methods, one-dimension (1D), block and two-dimension (2D) operations, and two precisions, single and double precisions. In the condition of single precision, the peak acceleration ratio of 1D, block and 2D operations is 220,240 and 77, respectively; in the condition of double precision, the peak acceleration ratio of 1D, block and 2D operations is 103,107 and 35, respectively. When the cluster size is small, the acceleration ratio of 2D operation reach overwhelming superiority over the rest operations. The acceleration ratio is very low at 1D and block operations. When the cluster size is medium,2D operation can’t complete the calculation. The acceleration ratio grows rapidly until reach the peak acceleration ratio and it is far more than that of 1D operation. When the cluster size is large, both 1D and block operations have reached the peak acceleration ratio. The peak acceleration ratio of block operation is slightly greater than that of 1D.2、Application of GPU acceleration in structural optimization of (Al2O3)n clusterWe use Genetic Algorithms-GPU (GA-GPU) method to predict and optimize the structures of (Al2O3)n (n=1-15) clusters and analysis their structural characteristics. Because of the small cluster size and the demand of high accuracy, we choose two-dimension operation in condition of double precision, the peak acceleration ratio is about 35. Conclusions are as follows:The global minimum structures are cage, cage and tea-cozy for n=1,2 and 3 respectively. At n= 4 and 5, it is highly symmetrical cage structure, disordered at n= 6. It is highly symmetrical large cage at n=7,8 and 9 and disordered at n= 10. We find that all the reported structures of (Al2O3)/, (n=1-10) clusters have been successfully located by our method. On the other hand, we also predict the structures of (Al2O3)n (n=11-15) clusters, which have never been reported in previous literatures. The structure is similar to the global minima of (Al2O3)7at n= 11. It is disordered at n= 12, but some atoms can overlap in the specific perspective. At n=13 and 14, the symmetry is Cs and D2, respectively, disordered at n=15. |
PDF Full Text Request