Assuming you compare apples to apples (IE put them both in the same precision modes, tell them both to actually follow the standard[1], etc), it's not appreciab

Assuming you compare apples to apples (IE put them both in the same precision modes, tell them both to actually follow the standard[1], etc), it's not appreciably faster for anything except really specialized kernels. [1] A lot of compilers take "liberties" with the standard by default in the name of performance. See for example, IBM's qstrict and qstrict_induction. The default at O2 and above is Qstrict and Qnostrict_induction (IE "change the semantics of my program"). So if you really want to compare performance numbers between compilers, you have to normalize them to something first, be it "go crazy and do what you want to my program" or "follow the standard" People who compare compiler benchmarks rarely actually do this. Note that the use cases of the compilers are very different, so it's probably reasonable for IBM/Intel/etc to have these as defaults and GCC/LLVM to have different defaults, but it doesn't change what you would want to compare if you want to compare generated code performance.