I like the version from @mar10, though from my point of view, there is a chance of misusage (seems it is not the case if versions are compatible with Semantic Versioning document, but may be the case if some "build number" is used):
versionCompare( '1.09', '1.1');  // returns 1, which is wrong:  1.09 < 1.1
versionCompare('1.702', '1.8');  // returns 1, which is wrong: 1.702 < 1.8
The problem here is that sub-numbers of version number are, in some cases, written with trailing zeroes cut out (at least as I recently see it while using different software), which is similar to rational part of a number, so:
5.17.2054 > 5.17.2
5.17.2 == 5.17.20 == 5.17.200 == ... 
5.17.2054 > 5.17.20
5.17.2054 > 5.17.200
5.17.2054 > 5.17.2000
5.17.2054 > 5.17.20000
5.17.2054 < 5.17.20001
5.17.2054 < 5.17.3
5.17.2054 < 5.17.30
The first (or both first and second) version sub-number, however, is always treated as an integer value it actually equals to. 
If you use this kind of versioning, you may change just a few lines in the example:
// replace this:
p1 = parseInt(v1parts[i], 10);
p2 = parseInt(v2parts[i], 10);
// with this:
p1 = i/* > 0 */ ? parseFloat('0.' + v1parts[i], 10) : parseInt(v1parts[i], 10);
p2 = i/* > 0 */ ? parseFloat('0.' + v2parts[i], 10) : parseInt(v2parts[i], 10);
So every sub-number except the first one will be compared as a float, so 09 and 1 will become 0.09 and 0.1 accordingly and compared properly this way. 2054 and 3 will become 0.2054 and 0.3.
The complete version then, is (credits to @mar10):
/** Compare two dotted version strings (like '10.2.3').
 * @returns {Integer} 0: v1 == v2, -1: v1 < v2, 1: v1 > v2
 */
function versionCompare(v1, v2) {
    var v1parts = ("" + v1).split("."),
        v2parts = ("" + v2).split("."),
        minLength = Math.min(v1parts.length, v2parts.length),
        p1, p2, i;
    // Compare tuple pair-by-pair. 
    for(i = 0; i < minLength; i++) {
        // Convert to integer if possible, because "8" > "10".
        p1 = i/* > 0 */ ? parseFloat('0.' + v1parts[i], 10) : parseInt(v1parts[i], 10);;
        p2 = i/* > 0 */ ? parseFloat('0.' + v2parts[i], 10) : parseInt(v2parts[i], 10);
        if (isNaN(p1)){ p1 = v1parts[i]; } 
        if (isNaN(p2)){ p2 = v2parts[i]; } 
        if (p1 == p2) {
            continue;
        }else if (p1 > p2) {
            return 1;
        }else if (p1 < p2) {
            return -1;
        }
        // one operand is NaN
        return NaN;
    }
    // The longer tuple is always considered 'greater'
    if (v1parts.length === v2parts.length) {
        return 0;
    }
    return (v1parts.length < v2parts.length) ? -1 : 1;
}
P.S. It is slower, but also possible to think on re-using the same comparing function operating the fact that the string is actually the array of characters:
 function cmp_ver(arr1, arr2) {
     // fill the tail of the array with smaller length with zeroes, to make both array have the same length
     while (min_arr.length < max_arr.length) {
         min_arr[min_arr.lentgh] = '0';
     }
     // compare every element in arr1 with corresponding element from arr2, 
     // but pass them into the same function, so string '2054' will act as
     // ['2','0','5','4'] and string '19', in this case, will become ['1', '9', '0', '0']
     for (i: 0 -> max_length) {
         var res = cmp_ver(arr1[i], arr2[i]);
         if (res !== 0) return res;
     }
 }