<p>A smoothed particle hydrodynamics (SPH) has obtained wildly application to granular flow and soil failure problems in last two decades. The computational efficiency is limited by the number of particles, which makes it difficult for SPH to be applied to large-scale examples. In this study, we develop a three-dimensional SPH model based on Drucker–Prager closure with a non-associated plastic flow rule, which is accelerated by employing the GPU technology. A typical three-dimensional granular slope case is simulated with 44 million particles for 88.5 hours. GPU acceleration technology significantly improves the computing efficiency almost 200 times than single-core CPU for large scale geotechnical problems with more than 10 million SPH particles. Multiple shear bands are observed in this simulation, which reveal the failure mechanism of granular flow.</p>