The problem with Microbenchmarks is that you **really** need to understand what it is yo uare measuring.

As someone has already pointed out, the latest Sun VM does SIMD in the compiler.  So if it sees a set of bytecodes that can be folded down to SIMD the result will be much faster then if it doesnt and does them with discrete operations.

In english, this means that doing things like matrix multiplys are MUCH faster then doing the same numebr of unrelated mults.

Jeff