On the Utility of Graphics Cards to Perform Massively Parallel Simulation of Advanced Monte Carlo Methods

• Activity Number: 207
• Type: Invited
• Date/Time: Monday, August 1, 2011 : 2:00 PM to 3:50 PM
• Sponsor: JCGS-Journal of Computational and Graphical Statistics
• Author(s): Anthony Lee*+ and Chris Holmes and Arnaud Doucet and Chris Yau and Mike Giles
• Companies: University of Oxford and University of Oxford and University of British Columbia and University of Oxford and University of Oxford
• Keywords: GPU ; stochastic simulation ; Monte Carlo ; MCMC ; Particle Filter

Abstract

For certain types of scientific algorithms, desktop graphics cards using graphical processing units (GPUs) offer the performance of cluster-based computing at a fraction of the cost. Moreover, GPUs are dedicated, low maintenance, energy-efficient devices that are becoming increasingly easy to program. In this talk we overview the class of statistical algorithms amenable to GPU computation. We then present a case study using advanced Monte Carlo algorithms including population-based MCMC, sequential Monte Carlo samplers and the particle filter. We demonstrate that GPUs can lead to substantial speedups ranging from 35 to 500 fold over conventional CPU single-threaded computation. This suggests that GPUs and other multi-core devices are likely to change the landscape of high performance statistical computing in the near future.