Interesting question! I was not aware about the mentioned huge difference in I/O-speed. But here I just want/must cite the documentation:

? MX files cannot be exchanged between different operating systems or versions of the Wolfram System.

? [WDX] Stores arbitrary Wolfram Language expressions in a serialized, platform-independent form.

So, at least for me the .mx format would never ever be an option if data are meant to be persistent.

But is is well-known that the documentation is incorrect. It just works cross-platform. The only thing which does work is to load a .mx file produced in a newer version of Mathematica into an oder one.

Thanks for the note @Sander Huisman. In my problem case with .wdx, the export from was a GNU/Linux x8664 (Intel(R) Xeon(R) CPU E5-2660 v3 @ 2.60GHz) and the import was to a GNU/Linux x8664 (Intel(R) Core(TM) i5-5300U CPU @ 2.30GHz) so both sides would have been using little-endian format. In addition, the version of Mathematica was the same on both systems.

The most efficient format is .mx:

test = RandomReal[{-10, 10}, {24000, 11000}];
AbsoluteTiming[Export["big.mx", test];]

needs 15 seconds ( depending also on your hard drive speed I guess). Loading it back in by

test=Import["big.mx"];

is pretty fast, too, less than a second. I never ever use wdx. Always .mx, which is also the fastest way known to me to collect larger results from parallel subkernels back to the master kernel ...