I haven't seen any speculation on floating point performance yet, but with an integrated vector processor and high memory bandwidth (6.4 GB/sec), it ought to be plenty fast.
The eight-stage pipeline should give a big boost to integer performance too.