Python, with gmpy2, 0m1.008s on my laptop

import gmpy2

n = 20000000

a = gmpy2.mpz(1)
b = gmpy2.mpz(0)
p = gmpy2.mpz(0)
q = gmpy2.mpz(1)

count = n

while count > 0:
    if count % 2 == 0:
        psq = p * p
        qsq = q * q
        twopq = 2 * p * q

        p = psq + qsq    # p -> (p * p) + (q * q)
        q = twopq + qsq  # q -> (2 * p * q) + (q * q)

        count //= 2
    else:
        bq = b * q
        aq = a * q
        ap = a * p
        bp = b * p

        a = bq + aq      # a -> (b * q) + (a * q)
        a += ap          #              + (a * p)

        b = bp + aq      # b -> (b * p) + (a * q)

        count -= 1

print(b)

Java: 8 seconds to compute, 18 seconds to write

public static BigInteger fibonacci1(int n) {
    if (n < 0) explode("non-negative please");
    short charPos = 32;
    boolean[] buf = new boolean[32];
    do {
        buf[--charPos] = (n & 1) == 1;
        n >>>= 1;
    } while (n != 0);
    BigInteger a = BigInteger.ZERO;
    BigInteger b = BigInteger.ONE;
    BigInteger temp;
    do {
        if (buf[charPos++]) {
            temp = b.multiply(b).add(a.multiply(a));
            b = b.multiply(a.shiftLeft(1).add(b));
            a = temp;
        } else {
            temp = b.multiply(b).add(a.multiply(a));
            a = a.multiply(b.shiftLeft(1).subtract(a));
            b = temp;
        }
    } while (charPos < 32);
    return a;
}

public static void main(String[] args) {
    BigInteger f;
    f = fibonacci1(20000000);
    // about 8 seconds
    System.out.println(f.toString());
    // about 18 seconds
}

COW

 MoO moO MoO mOo MOO OOM MMM moO moO
 MMM mOo mOo moO MMM mOo MMM moO moO
 MOO MOo mOo MoO moO moo mOo mOo moo

Moo! (Takes a while. Drink some milk...)

I implemented the matrix multiplication method (from sicp, http://sicp.org.ua/sicp/Exercise1-19) in SBCL but it takes about 30 seconds to finish. I ported it to C using GMP, and it returns the correct result in about 1.36 seconds on my machine. It's about as fast as boothby's answer.

#include <gmp.h>
#include <stdio.h>

int main()
{
  int n = 20000000;

  mpz_t a, b, p, q, psq, qsq, twopq, bq, aq, ap, bp;
  int count = n;

  mpz_init_set_si(a, 1);
  mpz_init_set_si(b, 0);
  mpz_init_set_si(p, 0);
  mpz_init_set_si(q, 1);
  mpz_init(psq);
  mpz_init(qsq);
  mpz_init(twopq);
  mpz_init(bq);
  mpz_init(aq);
  mpz_init(ap);
  mpz_init(bp);

  while(count > 0)
    {
      if ((count % 2) == 0)
        {
          mpz_mul(psq, p, p);
          mpz_mul(qsq, q, q);
          mpz_mul(twopq, p, q);
          mpz_mul_si(twopq, twopq, 2);

          mpz_add(p, psq, qsq);    // p -> (p * p) + (q * q)
          mpz_add(q, twopq, qsq);  // q -> (2 * p * q) + (q * q) 
          count/=2;
        }

      else
       {
          mpz_mul(bq, b, q);
          mpz_mul(aq, a, q);
          mpz_mul(ap, a, p);
          mpz_mul(bp, b, p);

          mpz_add(a, bq, aq);      // a -> (b * q) + (a * q)
          mpz_add(a, a, ap);       //              + (a * p)

          mpz_add(b, bp, aq);      // b -> (b * p) + (a * q)

          count--;
       }

    }

  gmp_printf("%Zd\n", b);
  return 0;
}

Ocaml, 0.856s on my laptop

Requires the zarith library. I used Big_int but it's dog slow compared to zarith. It took 10 minutes with the same code! Most of the time was spent printing the damn number (9½ minutes or so)!

module M = Map.Make
  (struct
    type t = int
    let compare = compare
   end)

let double b = Z.shift_left b 1
let ( +. ) b1 b2 = Z.add b1 b2
let ( *. ) b1 b2 = Z.mul b1 b2

let cache = ref M.empty 
let rec fib_log n =
  if n = 0
  then Z.zero
  else if n = 1
  then Z.one
  else if n mod 2 = 0
  then
    let f_n_half = fib_log_cached (n/2)
    and f_n_half_minus_one = fib_log_cached (n/2-1)
    in f_n_half *. (f_n_half +. double f_n_half_minus_one)
  else
    let f_n_half = fib_log_cached (n/2)
    and f_n_half_plus_one = fib_log_cached (n/2+1)
    in (f_n_half *. f_n_half) +.
    (f_n_half_plus_one *. f_n_half_plus_one)
and fib_log_cached n =
    try M.find n !cache
    with Not_found ->
      let res = fib_log n
      in cache := M.add n res !cache;
      res

let () =
  let res = fib_log 20_000_000 in
  Z.print res; print_newline ()

I can't believe how much a difference the library made!

Go

It's embarrasingly slow. On my computer it takes a little less than 3 minutes. It's only 120 recursive calls, though (after adding the cache). Note that this may use a lot of memory (like 1.4 GiB)!

package main

import (
    "math/big"
    "fmt"
)

var cache = make(map[int64] *big.Int)

func fib_log_cache(n int64) *big.Int {
    if res, ok := cache[n]; ok {
        return res
    }
    res := fib_log(n)
    cache[n] = res
    return res
}

func fib_log(n int64) *big.Int {
    if n <= 1 {
        return big.NewInt(n)
    }

    if n % 2 == 0 {
        f_n_half := fib_log_cache(n/2)
        f_n_half_minus_one := fib_log_cache(n/2-1)
        res := new(big.Int).Lsh(f_n_half_minus_one, 1)
        res.Add(f_n_half, res)
        res.Mul(f_n_half, res)
        return res
    }
    f_n_half := fib_log_cache(n/2)
    f_n_half_plus_one := fib_log_cache(n/2+1)
    res := new(big.Int).Mul(f_n_half_plus_one, f_n_half_plus_one)
    tmp := new(big.Int).Mul(f_n_half, f_n_half)
    res.Add(res, tmp)
    return res
}

func main() {
    fmt.Println(fib_log(20000000))
}

C, naive algorithm

Was curious, and I hadn't used gmp before... so:

#include <stdio.h>
#include <stdlib.h>
#include <gmp.h>

int main(int argc, char *argv[]){
    int n = (argc>1)?atoi(argv[1]):0;

    mpz_t temp,prev,result;
    mpz_init(temp);
    mpz_init_set_ui(prev, 0);
    mpz_init_set_ui(result, 1);

    for(int i = 2; i <= n; i++) {
        mpz_add(temp, result, prev);
        mpz_swap(temp, result);
        mpz_swap(temp, prev);
    }

    printf("fib(%d) = %s\n", n, mpz_get_str (NULL, 10, result));

    return 0;
}

fib(1 million) takes about 7secs... so this algorithm won't win the race.

Haskell

This is my own try, although I did not wrote the algorithm by myself. I rather copied it from haskell.org and adapted it to use Data.Vector with its famous stream fusion:

import Data.Vector as V
import Data.Bits

main :: IO ()
main = print $ fib 20000000

fib :: Int -> Integer
fib n = snd . V.foldl' fib' (1,0) . V.dropWhile not $ V.map (testBit n) $ V.enumFromStepN (s-1) (-1) s
    where
        s = bitSize n
        fib' (f,g) p
            | p         = (f*(f+2*g),ss)
            | otherwise = (ss,g*(2*f-g))
            where ss = f*f+g*g

This takes around 4.5 seconds when compiled with GHC 7.0.3 and the following flags:

ghc -O3 -fllvm fib.hs

Haskell

On my system, this runs almost as fast as FUZxxl's answer (~18 seconds instead of ~17 seconds).

main = print $ fst $ fib2 20000000

-- | fib2: Compute (fib n, fib (n+1)).
--
-- Having two adjacent Fibonacci numbers lets us
-- traverse up or down the series efficiently.
fib2 :: Int -> (Integer, Integer)

-- Guard against negative n.
fib2 n | n < 0 = error "fib2: negative index"

-- Start with a few base cases.
fib2 0 = (0, 1)
fib2 1 = (1, 1)
fib2 2 = (1, 2)
fib2 3 = (2, 3)

-- For larger numbers, derive fib2 n from fib2 (n `div` 2)
-- This takes advantage of the following identity:
--
--    fib(n) = fib(k)*fib(n-k-1) + fib(k+1)*fib(n-k)
--             where n > k
--               and k ≥ 0.
--
fib2 n =
    let (a, b) = fib2 (n `div` 2)
     in if even n
        then ((b-a)*a + a*b, a*a + b*b)
        else (a*a + b*b, a*b + b*(a+b))

C with GMP, 3.6s

Gods, but GMP makes code ugly. With a Karatsuba-style trick, I managed to cut down to 2 multiplies per doubling step. Now that I'm reading FUZxxl's solution, I'm not the first to have the idea. I've got a couple more tricks up my sleeve... maybe I'll try 'em out later on.

#include <gmp.h>
#include <stdio.h>

#define DBL mpz_mul_2exp(u,a,1);mpz_mul_2exp(v,b,1);mpz_add(u,u,b);mpz_sub(v,a,v);mpz_mul(b,u,b);mpz_mul(a,v,a);mpz_add(a,b,a);
#define ADD mpz_add(a,a,b);mpz_swap(a,b);

int main(){
    mpz_t a,b,u,v;
    mpz_init(a);mpz_set_ui(a,0);
    mpz_init(b);mpz_set_ui(b,1);
    mpz_init(u);
    mpz_init(v);

    DBL
    DBL
    DBL ADD
    DBL ADD
    DBL
    DBL
    DBL
    DBL ADD
    DBL
    DBL
    DBL ADD
    DBL
    DBL ADD
    DBL ADD
    DBL
    DBL ADD
    DBL
    DBL
    DBL
    DBL
    DBL
    DBL
    DBL
    DBL /*Comment this line out for F(10M)*/

    mpz_out_str(stdout,10,b);
    printf("\n");
}

Built with gcc -O3 m.c -o m -lgmp.

Sage

Hmm, you seem to assume that the fastest is going to be a compiled program. No binary for you!

print fibonacci(2000000)

On my machine, it takes 0.10 cpu seconds, 0.15 wall seconds.

edit: timed on the console, instead of the notebook

Mathematica, interpreted:

First@Timing[Fibonacci[2 10^6]]

Timed:

0.032 secs on my poor man's laptop.

And of course, no binary.