Case 1: Grapes Result Difference Analysis
A customer uses Grapes 3.2 to forecast the global weather in the 0.125 resolution in 10 days, and the number of grids is 2884 x 1441. After the Kunpeng cluster is deployed, the correlation coefficient between the calculation result of the Kunpeng cluster and that of the x86 cluster is 0.75, indicating that the two results are greatly different.
Based on the difference cause analysis, it is determined that the difference is caused by the -fp-model=fast compilation option enabled for x86 ICC. This option uses a performance optimization policy that sacrifices precision to achieve a faster calculation speed. After the compilation option -fp-model is changed to precise, the correlation coefficient between the Kunpeng calculation result and the x86 calculation result is 1, indicating that the results are the same.
