I tested this just to make sure . . . there was a savings of approx 2 seconds (21 vs 23)

Usually we like to see Benchmark code to back up this type of assertion. Using a the setup from PodMaster's (i.e. generating a random $string to test against), I get the following results, which seem to validate your results:

$ perl match
       Rate orig   lc
orig 1492/s   --  -7%
lc   1608/s   8%   --

$ perl match
       Rate orig   lc
orig 1506/s   --  -7%
lc   1620/s   8%   --

$ perl match
       Rate orig   lc
orig 1492/s   --  -8%
lc   1627/s   9%   --
[download]

And, in case you want to test this for yourself (or modify it to be more accurate to what you're doing), here's the code:

use strict;
use Benchmark qw(cmpthese);

my $pattern = '(new|old|number|start|simple|cross|heavy|die|exit)';

my @words = qw( new Old Number start Simple Cross heavy die Exit );
my $string = join ' 0\4/f ',
             map( { rand $_ } 1 .. 60),
             map { $words[ rand @words ] }
             1 .. 20;

cmpthese (-2, {
                orig => sub {    $string =~ /$pattern/i },
                lc   => sub { lc $string =~ /$pattern/  },
              }
         );
[download]