This paper presents a methodology to simulate and visualize crowds. Our goal is to represent the most realistic possible scenarios in a city. Due to the high demand of resources a GPU Cluster is used. We use real data from which we identify the behavior of the masses applying statistical and artificial intelligence techniques. In order to take advantge of the processing power of the GPU cluster we use the following programming models during the characters simulation: MPI, OmpSs and CUDA. We developed different visualization schemes: a) In situ, b) Streaming, c) Web. The web scheme is the most flexible, allowing to interact in real time with the simulation through a web browser. For this scheme we use WebGL and Cesium.