I don't know if this actually performs a valid comparison, but:

use warnings;
use strict;

use Benchmark qw(timethese);
use Module::Load;

BEGIN {
    timethese(
        1000000000,
        {
            my_use => sub {
                use Data::Dumper;
            },
            my_check => sub {
                load Data::Dumper if ! $Data::Dumper::VERSION;
            }
        }
    );
}
[download]

Output:

perl bench_use.pl 
Benchmark: timing 1000000000 iterations of my_check, my_use...
  my_check: 24 wallclock secs (27.21 usr + -0.09 sys = 27.12 CPU) @ 36
+873156.34/s (n=1000000000)
  my_use:  -3 wallclock secs ( 0.92 usr +  0.33 sys =  1.25 CPU) @ 800
+000000.00/s (n=1000000000)
[download]