Looking at other JVM languages (JRuby, Jython), it's at least possible that Rakudo-on-JVM could outperform Rakudo-on-Parrot.
However, it most likely won't be a silver bullet as far as performance goes, and even if it were, we'd be stuck with a perl6 with far different performance characteristics (startup time, memory consumption) than perl5.
On the positive side, we might get a well-defined VM abstraction layer out of it, which would be pretty valuable if someone rebooted a Parrot2 - and frankly, from what I can see, starting from scratch with a clearly defined goal might be easier than trying to backport Rakudo's abstraction layer into Parrot (among other things, Rakudo comes with its own object model, dispatch system and native call interface).