declspec usage is the same in ICL as in MSVC. Setting AVX option by itself doesn't increase data alignments. I see nothing in the docs to indicate that -Qsfalign may have been extended beyond 16-byte alignment, so you will need the declspec or equivalent if you want 32-byte alignment. You would only require 32-byte alignment if you specify syntax which requires it, such as #pragma vector aligned, __assume_aligned__, or aligned move intrinsics. Other AVX intrinsics support unaligned data (but 32-byte alignment may show performance advantage). You have a contradiction in terms between "official" release and beta. If you have the beta compiler, you have what documentation there was. It looks somewhat preliminary to me. Don't expect a .pdf. As Georg implied, the 2011 documentation should still be useful (and more "official") until we have a 2013 release.