To get the compiler and linker options for using the MKL you can use the Intel® Math Kernel Library Link Line Advisor

As on all non-intel chips, MKL will not automagically select the correct code path for fastest exectuion. The same holds true for multi-code-path (fat) binaries generated by the intel compilers (ifort/icc).

However, for MKL there is a trick to work around this problem by setting an undocumented environment variable: MKL_DEBUG_CPU_TYPE=N, where N is an integer that corresponds to a vectorization set. N=5 should force MKL to use AVX2 instructions, instead of the default SSE2 on non-intel chips.