Simulating complex physical phenomena implies the manipulation of an important amount of data. In order to simulate very large simulation domains on a limited computing architecture, such as industrial infrastructures, solutions have to be proposed. In this paper, a new out-of-core method is introduced in order to perform fast physical simulations using a complex Lattice Boltzmann model (LBM) on a single-node multi-GPU (CUDA) architecture. GPU global memory generally is far lower than the CPU main memory, can be problematic for a large simulation domain. The objective of this paper is to propose an efficient method of data exchanges between GPUs, the CPU main memory, which allows to perform fast complex simulations on large installations. The combination of this method with the massive parallelism of GPUs allows to keep good simulation performance. A complex simulation involving two physical components (water + air) is used in order to validate this method.