The purpose of this work is to make possible and speed up a grain system simulation implemented entirely on GPU, using the Discrete Element Method (DEM). The goal of implementing all the system on GPU is to avoid the cost of data transfer between the graphics hardware and the CPU. The proposed system simulate particles of different diameters, with collision treatment between particles and between particles and the environment. The Discrete Element Method consider normal forces and tangential forces applied on the particles. Parallel algorithms were designed to construct and storage the tangential forces history present in each contact between particles. Two ideas for the construction of the regular grid of cells are proposed and compared to perform the collision detection. The first one is very efficient to particles with fixed radius, while the second one shows more scalability in models with radius variation. The system consists of several algorithms running in threads, responsible for each step of the simulation. The results of the simulation were validated with the commercial program called PFC3D. The GPU particle system can be up to 10 times faster then the commercial program.