Your benchmark is unfair as the key extraction for ambrus solution is not being measured.

Anyway, using Sort::Key is, as usual, the fastest solution!

use Sort::Key 'ikeysort';

my @d = <DATA>; my @e;
use Benchmark qw/cmpthese/;
cmpthese (4e4, { grt    => sub { @e = map { substr $_, 6 }
                                      sort map { my $val = (split /\|/
+)[4]; sprintf "%06d$_",$val} @d; },
                 raw    => sub { @e = sort { (split '\|', $a, 12)[4] <
+=> (split '\|', $b, 12)[4] } @d; },
                 ambrus => sub {  my @key = map { (split /\|/)[4] } @d
+;
                                  @e = @d[ sort { $key[$a] <=> $key[$b
+] } 0 .. @d - 1 ]; },
                 sk     => sub { @e = ikeysort { (split '\|', $_, 12)[
+4] } @d }
               });
[download]

on my computer says...

         Rate    raw    grt ambrus     sk
raw    1887/s     --   -63%   -69%   -74%
grt    5141/s   172%     --   -15%   -29%
ambrus 6061/s   221%    18%     --   -16%
sk     7246/s   284%    41%    20%     --
[download]