Application of the OpenCL API for Implementation of the NIPALS Algorithm for Principal Component Analysis of Large Data Sets

An implementation of the nonlinear iterative partial least squares algorithm (NIPALS) was used as a test case for use of OpenCL for computation on a general purpose graphics processing unit (GPGPU) cluster using MPI. Timing results are shown along with results of a model of time required per iteration for defined problem sizes. Various steps in optimization of the code are discussed, moving from use of a single GPU, to multiple GPUs on a single node, to multiple GPUs on multiple nodes. Comparison of performance between OpenCL and BLAS implementations, modern CPU architectures and NVidia Tesla and Fermi class GPU systems are given.