Shuffle up and deal: accelerating GPGPU Monte Carlo simulation with application to option pricing

摘要

In this paper, we demonstrate some speedup opportunity regarding Monte Carlo simulation on graphic processingrnunit architecture, with financial application. We leverage on the possibility of reducing the volumernof actually generated random numbers, by replacing the generation phase with some shuffling using ComputernUnified Device Architecture’s built-in shuffle instructions. We will study various shuffling patterns andrnduration, elect the best among them with regard to induced correlation, using Granger causality test. Wernwill then study the accuracy and variance of results actually achieved by our general-purpose computing onrngraphic processing unit shuffled Monte-Carlo, exhibiting a computational time reduced by half while errorrnremains marginal.