With the upcoming SDK 7 (now available as a release candidate, installer here for registered developers), there is a new library of sparse and dense solvers that claim to be up to 6x faster than equivalent MKL functions. It also has solvers for eigen problems (yes, it's eigen, not Eigen!). This should be great for all kinds of applications that need to solve linear systems via Cholesky, LU, SVD, and QR decompositions. I love how NVIDIA keeps expanding their free GPU-powered libraries!
If you get excited about new versions and features like I do, have a look at these brief NVIDIA announcement videos:
- What's New in CUDA 7?
- What's New in Thrust 1.8? This is an awesome STL-like library with CUDA, OpenMP, TBB and CPU backends (I believe). New support for calling the Thrust API from kernels (device-side API), allowing sequential execution or dynamic parallelism.
- CUDA SDK 7 Release Notes (pdf)
