Even the US Department of Defense is getting into CUDA & GPGPU, and their Data Analysis & Assessment Center (DAAC) have begun posting a good set of very detailed tutorials on their website. The DAAC is a division of the DoD’s High Performance Computing Modernization Program (HPCMP), which specializing in maintaining and running supercomputers for various defense researchers.
It looks like they’re still in-progress, but it’s a great collection of very low-level tutorials showing exact translations between classic C & Fortran codes to their CUDA equivalents. They also have details on enabling double-precision support, working with CUDA error messages, and synchronization issues when using multiple threads.