But at the same time, you've just added a transcendental function to every loop iteration, which is a much more expensive operation than a multiplication. As I pointed out in Re^3: how to multiply an array of number as fast as posible. A more fair comparison would be scaling the operations so that we avoid INFs in the first place:

#!/usr/bin/perl
use strict;
use warnings;
use Benchmark qw':all :hireswallclock';
use threads;

cmpthese(10, {
    'for' => sub {
        my $k = 1;
        for (230000 .. 900000) {
            $k *= 1 + $_* 0.0000000001;
        }
        #print "for: $k\n";
        },
    'threads' => sub {
        my $thr2 = async {
            my $k = 1;
            for (565001 .. 900000) {
                $k *= 1 + $_* 0.0000000001;
            }
            return $k;
        };
        my $k = 1;
        for (230000 .. 565000) {
            $k *= 1 + $_* 0.0000000001;
        }
        $k *= $thr2->join();
        #print "threads: $k\n";
    }
});
[download]

I've replaced one multiplication per iteration with 2 multiplications and an addition (plus the store), which should shift the balance toward threads. And yet, the benchmark comes out in favor of single-threaded operation:

          Rate threads     for
threads 2.92/s      --    -35%
for     4.48/s     53%      --
[download]