The graphics processing unit (GPU) has emerged as a powerful and costeffective processor for general performance computing. GPUs are capable of anorder of magnitude more floating-point operations per second as compared tomodern central processing units (CPUs), and thus provide a great deal ofpromise for computationally intensive statistical applications. Fitting complexstatistical models with a large number of parameters and/or for large datasetsis often very computationally expensive. In this study, we focus on Gaussianprocess (GP) models -- statistical models commonly used for emulating expensivecomputer simulators. We demonstrate that the computational cost of implementingGP models can be significantly reduced by using a CPU+GPU heterogeneouscomputing system over an analogous implementation on a traditional computingsystem with no GPU acceleration. Our small study suggests that GP models arefertile ground for further implementation on CPU+GPU systems.
展开▼
