Physics-based fluid simulation has been widely used in manufacture and entertainment industry. However, numerically solving the three-dimensional incompressible fluid equation is an expensive computation task. Recently, the general purpose graphics hardware allows to speed up the numerical computations. To this end, we explore a new unified particle model on a multi-graphics processing unit (GPU) platform for interactive fluid simulation. In this paper, we present a parallel framework for fluid simulation with smoothed particle hydrodynamics. To begin with, we discuss the characteristics of storage dependence, data dependence, and decomposition method based on particles. And then, we design the computation model of simulation including kernel functions, adaptive time step, force, and pressure equations. The unified particle model not only ensures the mass and momentum conservation of simulation but also protects the volume of fluid and allows larger time steps, thus alleviating the calculation burden, contributing to the real-time simulation. Moreover, an effective parallel architecture based on the multi-GPU is implemented. For the fluid scene, the spatial domain is automatically partitioned into grid layers and assigned to multiple GPUs subsequently. A dynamic load balancing algorithm and an asynchronous data transferring strategy are proposed and implemented. Furthermore, we propose an adaptive density model for particle surface reconstruction. Finally, the quality and performance of the method are demonstrated using multiple scenes with different numbers of particle. Copyright © 2013 John Wiley & Sons, Ltd.