You can push directly onto the hash value,

use Data::Dumper;
my %hash;
my $sub_length = 2;

while (<DATA>) {
    chomp;
    for my $j (0 .. length - $sub_length) {
        push @{$hash{$j}}, substr $_, $j, $sub_length;
    }
}

print Dumper \%hash;

__DATA__
ABDC
EFGH
[download]

which does what you want. I've eliminated intermediate steps and temporary variables, and replaced the C-style loop with a more perly one. The code will work even if the data lines are not all the same length. Note that the length function is called on $_ by default.

Your code appears to overwrite the results from previous data lines each time. Is that the result you see?