Embracing GPU Technology in Bayesian Computation: Massively Parallel Computing on Your Desktop

• Activity Number: 5
• Type: Invited
• Date/Time: Sunday, August 1, 2010 : 2:00 PM to 3:50 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Marc Suchard*+
• Companies: University of California, Los Angeles
• Keywords: Bayesian ; Parallel computing ; Stochastic processes ; Optimization ; Bioinformatics ; mixture models

Abstract

Massive data present unique challenges to Bayesian statistics. I describe many-core computing algorithms that harness inexpensive graphics processing units (GPUs) to tackle this computation bottleneck. High-end GPUs contain hundreds of cores and are low-cost. I focus on efficient parallelization of integrating the complete data likelihood that plagues the statistical inference of many partially observed stochastic processes, including continuous-time Markov chain models in bioinformatics and Gaussian processes for spatial statistics. Novel algorithms demonstrate 100- to 200-fold speed-up on desktop computers. I conclude with a discussion of the future of many-core computing in statistics and touch upon recent experiences with statistical optimization problems and massively large and high-dimensional mixture models.