It's definitely just a standard lots-of-bits. Every bit of information is needed, lose one and the whole test will be wrong. For current wavefront exponents, like 2^60,000,000-1, save files are around 8 MB, or ~60M bits -- exactly what you'd expect.
Also Prime95 doesn't use GMP, its code is all hand-coded assembly, optimized specifically for the LL test and x86 architecture, written by one very dedicated George Woltman.
One last clarification: the reason to use Mlucas, and not Prime95 since it's so much faster, is for independent-code verification of the prime. (CUDALucas also counts in that regard, but the more independent triple and quadruple checks, the merrier!)
It's a myriad of answers. The Mlucas run used a larger FFT than was necessary, partly due to multithreading reasons. The Mlucas code also only uses SSE2 instructions, where the overclocked i7-2600K using GIMPS' Prime95 program is fully AVX capable. I would also expect that the i7 runs at a higher frequency, with a higher IPC, than the server. The "main" CUDALucas run took around 90 hours on a GTX 580. We didn't even hear about the 560 Ti run by Gilchrist, that was on the side.
Also Prime95 doesn't use GMP, its code is all hand-coded assembly, optimized specifically for the LL test and x86 architecture, written by one very dedicated George Woltman.