Not only that. They heavily rely on the massive vector unit (which, at the end, became AVX-512) which is integrated on each core, to squeeze the most of the theoretical computational power.
Another thing is that there are different Xeon Phi architectures, which aren't 100% compatible each other. I don't recall now after so long, but maybe KNL (Knights Landing) should be compatible with AVX-512.
Last but not really least, those are just coprocessors. Each core hasn't the full x86-64 instruction set, but some instructions are missing. That's they why they are sold as PCI cards: they need a regular PC to boot the o.s..
Since they are coprocessors, they require special code to be used: the PC's CPU should send data and code to the PCI card, and instruct the underlying card o.s. (there's a special Linux version there) to run the proper Xeon Phi code. After that the execution is done, the resulting data can be transferred to the PC's memory.
This is just a simplification to let you understand that things aren't easy with those coprocessors (albeit there are libraries and/or C/C++/Fortran compiler directives which makes the life quite easy).
TL;DR: don't use them.
P.S. I was a senior QA engineer at Intel working specifically for those coprocessors (and for the SGX extensions).
