Optimizing the NPB CG benchmark for multi-core AMD Opteron microprocessors

CG approximates the largest eigenvalue of a sparse, symmetric, positive definite matrix, using inverse iteration [3]. The matrix is generated by summing outer products of sparse vectors, with a fixed number of nonzero elements in each generating vector. The matrix sizes and total number of nonzero elements (“computed nonzeros,” following [3]) are listed in Table 1. The benchmark computes a given number of eigenvalue estimates, referred to as “outer iterations,” using 25 iterations of the conjugate gradient method to solve the linear system in each outer iteration.