Compiling CUDA applications typically requires you to pass the level of Compute Capability to the compiler. For example, Genga is compiled by typing

$ make SM=XX

where XX is the Compute Capability of the cards the code will run on. For cards available at the ICS, use the following.

You should compile on nodes of the partition you are running on. To do so, launch an interactive job as described on the SLURM Scheduler page.

Note that the GTX 680 cards have gimped double precision performance compared to the GTX 590. Do not be surprised if your code runs slower than you would expect.