According to the computer language benchmark game, the LuaJIT implementation seems to beat every other JIT-ed dynamic language (V8, Tracemonkey, PLT Scheme, Erlang HIPE) by an order of magnitude.
I know that these benchmarks are not representative (as they say: "Which programming language implementations have the fastest benchmark programs?"), but this is still really impressive.
In practice, is it really the case? Someone have tested that Lua implementation?
There's a good discussion at Lambda the Ultimate. LuaJIT is very good.
Many people have reported impressive speedups on lua-l. The speedups are most impressive for pure Lua code; the trace compiler is not as effective when there are lots of calls to C functions in loadable library modules.
I made an experiment with the lesson learned here: http://www.sampalib.org/luajit2.0_tunning.html Some data are not that valid anymore ( maxmcode=1024 is enough ), but luajit brings a robust improvement on a 600 lines of code pure Lua script (no C call to hit perfs...) that is not a large scale application nor an embedded use case but much more than the benchmarks.