Augmented Reality (AR) is the act of overlaying 3D virtual objects into a real-world scene. Using robust computer vision algorithms, it is possible to perform AR using only a single video camera. However, these algorithms are very computationally expensive, and most proposed systems have to sacrifice accuracy for speed.Graphics Processing Units (GPUs), originally designed to power graphics-intensive 3D video games and now commonplace on most gaming PCs, can also be used for general purpose computations. We developed a computer vision-based AR system accelerated by a single GPU, allowing robust feature detection and matching to be performed in every frame.We conducted performance evaluations in both indoor and outdoor environments, with parameters optimized for maximum possible accuracy of recovered poses. Our AR system achieves a stable 10-12 frames per second at 640480 resolution on a laptop.