IIRC Java Numerics was a dead end precisely because the required optimizations went against the 'write once, run anywhere' ideal. I have heard of specialized Java runtimes that can do interesting things to approximate BLAS/Atlas, but never seen them first hand.

I wanted to cite "Improving Java Performance Using Dynamic Method Migration", Lattanzi 2004, but the site hosting the paper isn't loading at the moment.

-- http://math.nist.gov/javanumerics/