with cuda 6 the NVPP libraries now include full blas and lots of sparse solver support as well. It would be really nice to write and debug cpu side with eigen, then pass a parameter to certain Eigen objects that they can use the cuda version in release.
http://developer.download.nvidia.com/co ... Report.pdf

Has there been discussion on this?

This should be already possible for BLAS: http://eigen.tuxfamily.org/dox/TopicUsingIntelMKL.html