This work presents a new proposal for the implementation of a GPU-based particle system. The simulation runs entirely on the graphic processor, thus eliminating data transfer between the CPU and the GPU. The proposed system is able to simulate particles with different diameters in confined environments, including support for inter-particle collisions, constraints, and particle-obstacle collisions. Inter-particle collision detection is accomplished by subdividing the space into a regular grid of cells. On modern graphics cards, the system is able to simulate up to one million particles at interactive rate. It is also proposed a flexible approach for modeling the obstacles that define the environment, allowing the creation of different scenes without relying on shader re-coding. The system is divided in different shaders responsible for each stage of the simulation. One fragment program is responsible to advance the particles in time. After that a vertex program builds the space subdivision structure. The following stages (collision detection and response, and constraint solving) are performed only by fragment programs using the relaxation method.