Fast Conjugate Gradients with Multiple GPUs Export

@incollection{citeulike:6072691, abstract = {The limiting factor for efficiency of sparse linear solvers is the memory bandwidth. In this work, we describe a fast Conjugate Gradient solver for unstructured problems, which runs on multiple GPUs installed on a single mainboard. The solver achieves double precision accuracy with single precision GPUs, using a mixed precision iterative refinement algorithm. To achieve high computation speed, we propose a fast sparse matrix-vector multiplication algorithm, which is the core operation of iterative solvers. The proposed multiplication algorithm efficiently utilizes GPU resources via caching, coalesced memory accesses and load balance between running threads. Experiments on wide range of matrices show that our matrix-vector multiplication algorithm achieves up to 11.6 Gflops on single GeForce 8800 GTS card and CG implementation achieves up to 24.6 Gflops with four GPUs.}, author = {Cevahir, Ali and Nukada, Akira and Matsuoka, Satoshi}, citeulike-article-id = {6072691}, citeulike-linkout-0 = {http://dx.doi.org/10.1007/978-3-642-01970-8_90}, citeulike-linkout-1 = {http://www.springerlink.com/content/9m742203qp7802m7}, doi = {10.1007/978-3-642-01970-8_90}, journal = {Computational Science – ICCS 2009}, keywords = {2009, conjugate-gradient, csr, cuda, jds, linear-algebra, matrix, matrix-vector, multiplication}, pages = {893--903}, posted-at = {2009-11-05 08:58:04}, priority = {0}, title = {Fast Conjugate Gradients with Multiple GPUs}, url = {http://dx.doi.org/10.1007/978-3-642-01970-8_90}, year = {2009} }

TY - CHAP ID - citeulike:6072691 L3 - citeulike-article-id:6072691 N2 - The limiting factor for efficiency of sparse linear solvers is the memory bandwidth. In this work, we describe a fast Conjugate Gradient solver for unstructured problems, which runs on multiple GPUs installed on a single mainboard. The solver achieves double precision accuracy with single precision GPUs, using a mixed precision iterative refinement algorithm. To achieve high computation speed, we propose a fast sparse matrix-vector multiplication algorithm, which is the core operation of iterative solvers. The proposed multiplication algorithm efficiently utilizes GPU resources via caching, coalesced memory accesses and load balance between running threads. Experiments on wide range of matrices show that our matrix-vector multiplication algorithm achieves up to 11.6 Gflops on single GeForce 8800 GTS card and CG implementation achieves up to 24.6 Gflops with four GPUs. JF - Computational Science – ICCS 2009 EP - 903 TI - Fast Conjugate Gradients with Multiple GPUs SP - 893 KW - 2009 KW - conjugate-gradient KW - csr KW - cuda KW - jds KW - linear-algebra KW - matrix KW - matrix-vector KW - multiplication AU - Cevahir, Ali AU - Nukada, Akira AU - Matsuoka, Satoshi PY - 2009/// UR - http://dx.doi.org/10.1007/978-3-642-01970-8_90 ER -