Using common graphics hardware for multi-agent traffic simulation with CUDA

Today's graphics processing units (GPU) have tremendous resources when it comes to raw computing power. The simulation of large groups of agents in transport simulation has a huge demand of computation time. Therefore it seems reasonable to try to harvest this computing power for traffic simulation. Unfortunately simulating a network of traffic is inherently connected with random memory access. This is not a domain that the SIMD (single instruction, multiple data) architecture of GPUs is known to work well with. In this paper the authors will try to achieve a speedup by computing multi-agent traffic simulations on the graphics device using NVIDIAs CUDA framework.