Performance of single-processor BLAS on IBMp690

Abstract The Basic Linear Algebra Subprograms, BLAS, are the basic computa-tional kernels in most applications. BLAS 1 and BLAS 2, the vector-vector and matrix-vector routines, require memory accesses in the same order ascomputations and thus cannot achieve performance close to peak performance on modern computer architectures. BLAS 3 matrix-matrix operationson n * n-matrices on the other side can do order n3 operations with only or-der n2 memory accesses. This much better ratio of computation to memoryaccess allows for much higher performance. To show which performance can be expected using the BLAS routines from IBM's ESSL on an IBM p690 weinvestigated the performance of one routine of each BLAS level and compared it to that of the corresponding routines on a CRAY T3E. i