I ran the tests on my opteron server (2.2 GHz, RHEL 4, gcc-4.1.2) after compiling mzscheme 370 on the box. Previously I used the prebuilt (presumably 32-bit) mzscheme binaries and native 64-bit gambit binaries on MacOS, but this build uses 64-bit operations in both the mzscheme and gambit libraries, and mzscheme looks better here. If mzscheme were to use the fft multiply routines in gmp it would look better still; with the current binaries of gsi/gsc and mzscheme, mzscheme would look worse for larger numbers.
Anyway, here are the results:
370 b22
(expt 3 1000000) ; a 124 89 (expt 3 1000001) ; b 126 89 (* a a) ; c 221 106 (* a b) ; 221 150 (quotient c a) ; 560 687 (sqrt c) ; 402 667 (gcd a b) ; note 1 1959 2708 (* a b) 62 62 (expt1 3 1000000) ; note 2 124 85 (expt2 3 1000000) ; note 3 362 370 (* a a) ; a=3^1000000 221 105 (expt 10 10000000); a 228 193 (fib 10000) ; note 4 33 29 (factorial 10000) ; note 5 137 158 (partial-factorial 0 10000) ; note 6 9 13 (binary-splitting-compute-e 10000) ; note 7 539 968 (binary-splitting-compute-pi 10000) ; note 9 716 1220 (pi-brent-salamin) ; n. 10, beta^k=10^100000 2646 3939 (pi-brent-salamin) ; beta^k=2^332193 1486 2889