I took a sample of 10m perlin noise values, sorted them, and recorded the threshold value for each percent of total samples.
There will be some variation no matter what (that's why 50% isn't at exactly 0.5), but statistically speaking it should be close enough to not have any noticable difference in practice.