In fact, both were needed to optimize things. Mzscheme's doc says explicitly (though I can't remeber where exactly) that standalones are a little faster than interpreted code (there are a few more optimisations they can perform during compilation I guess). Running arc1.scm through mzscheme -f is a little slower.
There is an initial overhead for byte-compiling the code
and jitting it -- but that's not something that you'd be
able to measure for such a small piece of code. Once
that's done, it's the same code -- mzscheme (since a good
while ago) on intel and ppc does not interpret code. Ever.
Even on solaris, where the jit is disabled, it's
"interpreting" byte-compiled code, so it is not an
interpreter in any case.
