Implementation of probabilistic forecasts in a GPU for big data problems

High Performance Computing via General Purpose Graphical Processing Unit (GPU) is a potential instrument to speed up computational times. In a world where big data is becoming a revolution, GPU could play an important role. This work intends to analyze the performance of GPU by implementing the calculation of probabilistic forecasts based on single exponential smoothing in conjunction with simulated predictive distributions. Essentially, supply chain companies must deal with a high number of forecasts at SKU level. In this context, reducing the computational times can be a source of a competitive advantage. Since the forecasts are usually made independently between SKUs, this problem can be easily parallelized and GPU computing can exploit such parallelization. To the best of authors knowledge, this is the first time GPU is applied to a supply chain demand forecasting context. Firstly, we will show how to adapt the programming of probabilistic forecasts in a parallel fashion. Then, real data coming from a manufacturer company will be used to illustrate the differences between GPU and traditional CPU computing. The results show that GPU can significantly increase the computational speedup ratio more than 30 times with respect to traditional CPU computing