> Why should FP be slower ?
It's easier to move an object 1 cm to the left than it is to bootstrap a universe where the object happens to 1 cm to the left.
> Why can't the runtime automatically leverage multiple cores and speed up the process ?
That is more related to proper dataflow analysis than immutability, and this is much more mature in VM runtimes that are actually used (hundreds of man-centuries vs. the odd PhD thesis).