Did you try GCC with -O3 and -ffast-math? It seems like ICC has its analogue to ffast-math enabled by default: http://scicomp.stackexchange.com/questions/20815/performance...