Though discrete particle simulation (DPS) has been widely used for investigating gas-solid flows from a more detailed level as compared to traditional two-fluid models (TFMs), it is still seriously limited by the computational cost when large scale systems are simulated. GPUs (graphic processing units), with their massive parallel architecture and high floating point performance, provide new possibilities for large-scale DPS. In this paper, a multi-scale CPU (central processing unit)-GPU hybrid computation mode is developed, in which the fluid flow is computed by CPU(s) while the particle motion is computed by CPU(s). To explore its feasibility, this mode is adopted to simulate the flow structures in the fluidization of Geldart D and A particles, respectively. Further coupled with an EMMS (energy minimization multi-scale) based meso-scale model, the flow behavior in an industrial fluidized bed is finally simulated, shedding light on the engineering applications of DPS, as an alternative to TFM. (C) 2012 Elsevier B.V. All rights reserved.