The NVIDIA CUDA Sparse Matrix library (cuSPARSE) provides a collection of basic linear algebra subroutines used for sparse matrices that delivers up to 8x faster performance than the latest MKL. The cuSPARSE library is designed to be called from C or C++, and the latest release includes a sparse triangular solver.
cuSPARSE is >6x Faster Than Intel MKL
NVIDIA cuSPARSE performance on a standard battery of sparse matrix test cases
Up To 40x Faster Than Intel MKL with 6 CSR Vectors
cuSPARSE Tri-diagonal solver up to 14x Faster Than Intel MKL
The cuSPARSE library is freely available as part of the CUDA Toolkit at www.nvidia.com/getcuda.
You may also be interested in the CUSP library of C++ sparse matrix operations and graph algorithms, including sparse iterative solvers and several preconditioners. CUSP is an ongoing NVIDIA Research project available as an open source on Google Code.
For more information on cuSPARSE and other CUDA math libraries: