GPGPU : Metropolis hastings sampling on a cluster of GPUs
under guidance of Prof. Sharat Chandran
Recent advances in the speed and programmability of graphics hardware have made it a
compelling platform for computationally demanding tasks in a wide variety of application
domains. Many of the Bayesian machine learning algorithms rely on quantities estimated
over samples and quite often sampling is the single most computationally demanding task
in such learning methods. Fortunately, sampling techniques are fairly general and are easily
extended to new models. Sampling methods like Gibbs and MetropolisHastings
are easily parallelizable which make the sampling a perfect application for distributed
computing. Furthermore, their numerical simplicity makes them an ideal candidate for
implementation on GPUs. This project is about mapping Metropolis Hastings sampling algorithm to graphics hardware.
