Due to the expensive cost of many computer and physical experiments, it is important to carefully choose a small number of experimental points uniformly spreading out the experimental domain in order to obtain most information from these few runs. Although space-filling Latin hypercube designs (LHDs) are popu- lar ones that meet the need, LHDs need to be optimized to have the space-filling property. As the number of design points or variables becomes large, the to- tal number of LHDs grows exponentially. The huge number of feasible points makes this a difficult discrete optimization problem. In order to search the opti- mal LHDs efficiently, we propose a population based algorithm which is adapted from the standard particle swarm optimization (PSO) and customized for LHD. Moreover, we accelerate the adapted PSO for LHD (LaPSO) via graphic process- ing unit (GPU). According to the examined cases, the proposed LaPSO is more stable compared to other two methods and capable of improving some known results.