The VFX Forth compiler is about 3 to 3.5 times faster on most benchmarks. GForth-fast does some clever "JIT" type tricks to make it faster than a naive Forth system but there is still some overhead inherent in reading the lists of addresses in threaded code.
Right, the Gforth project makes use of lots of clever techniques, but they're not interested in conventional optimising native-code compilation (whether JIT or ahead-of-time) and if I understand correctly they're not interested in writing platform-specific code (i.e. hand-optimised assembly).
On another note, Gforth is actively being worked on, but it's been almost 7 years since the last stable release.