The new thing is that performance has improved, from 2x slower than native to 1.5x slower than native, and that a big chunk of that is due to float32 optimizations.