Ahh... This is another topic I wanted to start, but I was too lazy. There are obviously some big changes in this new algo in ways I still don't quite understand. It seems that the algo actually favors mid-range and small miners, because stronger machines don't bring proportionally as much HPS2 (real HPS measured by pool). Now it's much better to have 10 small machines which have 8 threads each than to have 1 big machine which has 80 threads.

Here is a rough comparison of my miners:

blu1    = 16 cores, genproclimit 32, 64k HPS, 42k HPS2
blu2-3 = 24 cores, genproclimit 48, 64k HPS, 42k HPS2
blu4    = 40 cores, genproclimit 80, 55k HPS, 27k HPS2
blu5-9 =   2 cores, genproclimit   8, 17k HPS, 35k HPS2

First of all, why do blu2 and 3 have the same hashrate as blu1? In the previous algo they had about 530k while blu1 had about 350k, and these numbers are exactly proportionate to the number of cores (24 cores = 530k hps, 16 cores = 350k hps, divide them and you'll see, about 22k per core). Now both the 16 core one and the 24 core ones have the same hashrate, as if the CPU on the 24 core machines is not used to its fullest, but it is, it has 100% usage.

But the real mystery here is how in the world does a $20/mo shared VPS (blu5-9) outperform a $250/mo dedicated enterprise machine (blu4)? In the previous algo the high-spec machine was at the top of the leaderboard at around 830k hashrate and the cheap VPSs had around 70k each.