Atmospheric transport and dispersion (T&D) models play an important roll in United States national defense. Due to operational time constraints, less sophisticated models have consistently dominated the defense market. Recent advances in graphics processing units (GPUs) and their programming models have made GPUs an attractive platform for commodity, low-power, high-performance parallel computing. Two GPU accelerated (using NVIDIA Corporation's CUDA technology) versions of a sophisticated, large-eddy simulation (LES) based, Lagrangian stochastic model, developed at the National Center for Atmospheric Research (NCAR), were implemented and compared against their single and multiple core CPU (Intel Harpertown) counterparts. The implementation representing the shortest route to GPU acceleration observed a single GPU speedup of 14x over the single core CPU implementation. A more robust and scalable single GPU implementation observed speedups of 20x over the single core CPU implementation.