I've been working with CUDA now for about a year now. Last summer I went out and bought a relatively cheap 8800 GT card, which even as an older card is capable of something like 500 GFlops if fully utilized. My Core 2 Duo can do about 20, and that's assuming you can code your programs with SIMD instructions.
Its not easy to utilise the full 500 GFlops, and this is for single FP precision, double FP precision only exists on the newer 200 series and throughput is much lower.
Actually, I think that number may be for integer instructions, which are the fastest (then single-precision, then double-precision). Which makes it even better for this scenario, since the hashing algorithm only uses integer ops.
