I tried that too, but it wasn't any faster – you've already gotten most of the gains to be had by lifting the random number generation out of the innermost loop. So I went with the option that uses less memory.