As T-Bone said, it's translated just once and the resulting translated code is stored in memory until it's needed again (like for subsequent calls to the same function). So next time that function is called, it'll be much faster.
Comparing interpreted code with JIT, a JIT is usually about ten times as fast. Certainly in the Kaffe Java virtual machine, the JIT is ten times as fast as the interpreter. What would be good then would be to write the translated code out to a sharable Amiga-like library, so that it doesn't have to be translated again. I don't think any one does this yet.
tony