Stochastic Speculative Computation Method on General Purpose Graphics Processing Units

Parallel processing is considered effective in order to solve problems with significant computational complexity. The development of graphics processing units (GPU) in recent years has led to eneral purpose GPU (GPGPU) and brought significant benefits to the AI research field. In conventional parallel processing, processes must be frequently synchronized; thus, parallel computing does not necessarily improve the efficiency of computation except for special algorithms designed for parallel computation. In this study, we investigate the effect of applying MultiStart based speculative computation on GPGPU. This method incurs little synchronization overhead. Although the effect of this method is stochastic, an expected value is theoretically calculable. We analyze theoretically about the effects of the speculative method, and provide the results of applying the method to combinatorial optimization problems.