visual computing for fields like physics, geology, meteorology, medical, etc are the main applications where you will see hybridization of HPC clusters involving fermi-like nodes.
i recall a demonstration of Dt MRI images
like this taking weeks to render in an opteron farm. you have to take eigenvalues from measurements of the magnetic field vector due to currents flowing along nerve axons in the brain in order to reconstruct the path of the current. i have no idea how this is done but you can see how the work can be resolved into SIMD-like vector components.
GPGPU can make a ton of money in scientific communities if the ease of developing in CUDA continues to increase.