Implementing the conjugate gradient algorithm on multi-core systems

In linear solvers, like the conjugate gradient algorithm, sparse-matrix vector multiplication is an important kernel. Due to the sparseness of the matrices, the solver runs relatively slow. For digital optical tomography (DOT), a large set of linear equations have to be solved which currently takes in the order of hours on desktop computers. Our goal was to speed up the conjugate gradient solver. In this paper we present the results of applying multiple optimization techniques and exploiting multi-core solutions offered by two recently introduced architectures: Intel's Woodcrest general purpose processor and NVIDIA's G80 graphical processing unit. Using these techniques for these architectures, a speedup of a factor three has been achieved.