It seems to optimise much more aggressively than clang or gcc.
The other thing worthy of note is that icc seems to be far less aggressive at exploiting undefined behaviour than e.g. clang (which it is extremely notorious for), contradicting their notion/theory that exploiting UB is necessary for good optimisation. From what I've seen, icc's main strength is in instruction selection and scheduling.
Optimizations based on language semantics (clang) and instruction selection (icc) should be quite complementary, so we could have a best-of-both compiler.
The other thing worthy of note is that icc seems to be far less aggressive at exploiting undefined behaviour than e.g. clang (which it is extremely notorious for), contradicting their notion/theory that exploiting UB is necessary for good optimisation. From what I've seen, icc's main strength is in instruction selection and scheduling.