There are a number of assumptions involved in benchmarking this, but my try shows s/// as twice as fast:

use Benchmark 'cmpthese';
use strict;
use warnings;
my $big;
$big .= join '',map chr, 0..255 for 0..255;
print length($big), " characters.\n";
sub subst { my $tmp; ($tmp=$big) =~ s/(.)/ord($1).' '/seg; $tmp }
sub unpac { my $tmp; $tmp = join ' ', unpack 'C*', $big; $tmp }

print length(subst()), " characters in ascii numbers.\n";
print "whoops!\n" if subst() ne unpac().' ';

cmpthese( -10, {
    subst => \&subst,
    unpac => \&unpac
});
[download]