My understanding (but I'm not an ICC user) is the majority of the performance improvement are from the vectorized math libraries (that technically you can use from GCC as well).

It uses a default lower accuracy floating point model than GCC, so initially people think it's faster til they realise their results are slightly different and have to mess around with the `fp-model` flag

Right, of course, IIRC it does the equivalent of -ffast-math by default.