Technically, it does not produce very bad code when run on AMD systems, it produces very bad code for running on AMD systems.

It avoids using some performance instructions which it uses on the Intel build and which would work identically on AMD.

Their excuse is that they can't guarantee they would work identically on AMD and so resort to much slower un-pipelined code. It's like Ferrari testing a Porsche and only using 1st gear because they can't be sure that other gears would perform in the same way

There is some truth to that, as it does not do all optimizations that it could; however, in my experience, even on AMD systems, ICC beats MSVC hands down.