Variance, if expressed as a percentile, should remain within the same limits regardless of overall computational power.
So when we hit 2PH/s, the variance could be seen as swinging hundreds of TH/s in a 3 day window?The estimated network hashrate is
(number of blocks found in time window) * (average number of hashes required to find a block) / (length of time window)
Block finding is always a Poisson process with rate of about 1 per 10 minutes, so the number of blocks will always follow the same distribution (Poisson distribution with mean (time window / 10 min)). If you increase the number of hashes per block, both the mean and the standard deviation scale by the same factor, so you have the same relative variance.
On the other hand, if instead you estimated using the difficulty-1 shares found, you would indeed have less variance, as there are more of the individual items you're tracking. But if you track blocks, and the difficulty of finding blocks scales to maintain a given rate, it doesn't matter.
