The results in general are much better:
time ./fannkuch -:m20000 11 > output_fannkuch
time ./fasta -:m20000 25000000 > output_fasta
time ./binary-trees -:m20000 16 > output-binary-trees
time ./pidigits -:m20000 2500 > output
A range of improvement is apparent, with binary-trees benefitting a lot and pidigits not really getting much better.
I picked a 20 meg heap because a) it seems wasteful for individual programs to reserve (and probably waste) huge chunks of memory, if you're running a lot of programs b) when I tried higher values up to 50 megs on the fasta benchmark, it was faster, but only by another .2 seconds or so; about 1-2%... and c) Most other languages, and in this case SBCL, don't seem to request more than about 20 megs for their initial heap size either.
Of note is the fact that the prebuilt version produced slightly faster code than the version I compiled myself, specifying the optimisations mentioned in the INSTALL.txt document. Not sure why... These numbers are for the Gambit I compiled myself, and I didn't feel like dumping it and re-installing the prebuilt version just yet.
Anyway, these are nice results!