We assess the performance of the combined Open Accelerator (OpenACC) programming standard and graphics processing unit (GPU) accelerator for lattice Boltzmann (LB) simulations of fluid flow, heat and mass transfer. By optimizing the data layout, minimizing the memory access frequency, and adjusting the number of gangs and vector length, we show that the enhanced parallel computations can result in orders of magnitudes of speedup relative to the serial implementation of the LB algorithm. Based on such implementations, benchmark quality results are obtained with fine grid of 20492 for both two-dimensional lid driven cavity flow with Reynolds number up to 7500, and double diffusive cavity flow with solute Rayleigh number up to 10(8). (C) 2017 Elsevier Ltd. All rights reserved.