Add this to your considerations. It does 600_000 char sequence in 3.39 seconds.

Updated to correct caveat.

#! perl -slw
use strict;
use List::Util qw[sum];
use Time::HiRes qw[ time ];

my $sequence = join'', map{ ( qw[ A C G T ] )[ rand 4 ] } 1 .. 600_000
+;

my $start = time;
my( %one, %two, %three );
$one{ $_ }++ for unpack '(A1)*', $sequence;
$two{ $_ }++ for unpack '(A2X)*', $sequence;
delete @two{ (qw'A C G T') };
$three{ $_ }++ for unpack '(A3XX)*', $sequence;
delete @three{ (qw'AA AC AG AT CA CC CG CT GA GC GG GT TA TC TG TT') }
+;

printf "Elapsed: %3f seconds\n", time - $start;
print sum values %one;
print sum values %two;
print sum values %three;

<STDIN>;
print "$_ => $one{ $_ }" for keys %one;;
print "$_ => $two{ $_ }" for keys %two;;
print "$_ => $three{ $_ }" for keys %three;;

__END__
P:\test>499988
Elapsed: 3.390625 seconds
600000
599999
599998
[download]

It has one caveat. Where the sequence is not a multiple of 2 or 3 characters, the last 1 or two chars will get counted. This is easily filtered.