it has been recognized in the last 5 years that specialized graphics hardware can also be used for general purpose computations. the architecture of these cards is such that simd computations are naturally a good fit for a certain class of applications. i will outline the programming model of modern graphics cards, sketch the history of the development of the supporting software stack and if there is interest i will outline how we have implemented lattice gauge theory algorithms leading to dramatic speedup of monte carlo simulations.