GPUMatrix is an Eigen-compatible GPU Matrix library: http://code.google.com/p/gpumatrix/

I'm wondering if anyone has experience with it and/or opinions about it?

On a related note, are any Eigen users also using CUBLAS? How do you do it? I'd prefer to have some compatibility and code reuse across CPU/GPU. By the way, the upcoming CUDA 6.0 is going to have "unified memory".

Check this out: https://github.com/bravegag/eigen-magma

Yes I'm looking forward to test the overhead of CUDA 6's unified memory. If that's not too much, this might simplify a lot our job to support a DeviceMatrix class.

Thanks, eigen-magma looks interesting, but it's rather unclear from looking at that github page what the status of the project is.

Also, does it use CUBLAS strictly as a BLAS backend, or does it allow you to keep the temporary results in the device memory?

AFAIK results are not kept in device memory.