AMD 955 + NV_CUDA faster than Intel i7 ?

Superior floating performance has been reported for the INTEL i7 over the AMD 9xx CPU. Can this <FP> performance deficit be made up by CUDA_programming with newer NVidia graphics cards ?

I'm trying to settle on homebuilt kit specialized for <FP> scientific calculations < NO GAMING >. The budget is tight ( ~$1-K ) . Combination i7/x58 INTEL kit usually runs $150-200 more than PH_9xx/am3 AMD kit while graphic cards are a "push".

Sure I can roll-my-own C-code. I'd like to hear from somebody who does it already. Is it practical in short_term to expect NV_CUDA graphics card function to make up the floating_point performance difference ... and in-fact overwhelm CPU <FP> performance. I've seen factors-of-10+ bandied about!