Band selection is a common technique to reduce the dimensionality of hyperspectral imagery. When the desired object information is known, the reduction process can be achieved by selecting the bands that contain the most object information. It is expected that these selected bands can offer an overall satisfactory detection and classification performance. In this paper, we propose a new particle swarm optimization (PSO) based supervised band-selection algorithm that uses the known class signatures only without examining the original bands or the need of class training samples. Thus, this method requires much less computing time than other traditional methods. However, the PSO process itself may introduce additional computation cost. To tackle this problem, we propose parallel implementations via emerging general-purpose graphics processing units that can provide satisfactory results in speedup when compared to the cluster-based parallel implementation.