async calculation of Fibonacci slower than sequence calculation

I'm trying this code for calculation of Fibonacci series:

unsigned long fib_async(unsigned long n) {
    if (n<2) {
        return n;
    }
    else {
        auto res1 = async(fib_async, n-1);
        auto res2 = async(fib_async, n-2);

        return res1.get() + res2.get();
    }
}

Why is this code slower than classical:

unsigned long fib(unsigned long n) {
    return n<2 ? n : fib(n-2) + fib(n-1);
}

?

Synchronous version for n=40 lasts 0.7 seconds on my 8-core i7 CPU. Asynchronous example for n=40 lasts more than 1 minute and it seems to run on one CPU core only.

What am I doing wrong? I'm testing it on a Linux machine with gcc 4.8.4.