I'm unsure of the value of benchmarks of hash function performance when you aren't defining what you need the hash function for. Python's built-in hash function is not designed to do the same things that the SHA family is.
The timings for hash() are meaningless, since Python's str object caches the hash after the first call [1]. FNV is one of the fastest hash functions, it should be at least suspicious that hash() is 500 times faster than FNV.<p>[1] <a href="http://hg.python.org/releasing/2.7.3/file/7bb96963d067/Objects/stringobject.c#l1266" rel="nofollow">http://hg.python.org/releasing/2.7.3/file/7bb96963d067/Objec...</a>
MD5/ SHA is designed to be slow. so it makes sense that one that is optimized to be chaotic, low collision probability and even distribution could be faster.