It's closer to 40% slower or a 1070 is 40% faster. That's based off raw CUDA cores. Just off the top, this holds true in Skunk. A 1060 3GB is producing 19 while top results for a 1070 are 32. Which is almost exactly 40% increase. That's from Yiimp. Half as fast was a typo, I meant 50% (but it's closer to 40%).
19 x 1.5 is 28.5. The 40% slower is straight from the difference in cuda cores.
1080/ti both have issues with algos due to their GDDR5X memory. Even algos that they work 'properly' in they don't always get the full speed. You can look at Epsylons thread for more information on that. I had to hammer out a couple really big posts about it and rather not rehash that here. Either way the 1080/ti are bad buys, especially the TIs which have much higher power draw and coolers generally can't keep up with them, in addition to the increased system power draw that requires much more beefy PSUs. They're practically worthless in Dagger. For instance a 1080ti should routinely be 85% faster then a 1070 for the amount of CUDA cores it has, but rarely is and often is closer to 50% depending on the algo.
Best buy still is a 1070, even at the slightly higher prices they have currently.
1) Skunk 1060 6GB ~21, 1070 ~31 1080 ~41
21*1.5 = 31.5
21*2 =42
Straight forward.
2) 1080/ti have issues only with dagger-hashimoto algo. I don't know other. Other algos are like skunk.
3) Your maths is completely wrong
1060 - 115W 21Mh = 5.5W/Mh
1080 - 165W 41Mh = 4W/Mh
Of course you can tweak both of them gaining more or less, but anyway 1080 is superior in w/MH
4) 1080/ti are practically worthless only in 1 algo, but thanks to Claymore they are good in Dual mining leaving 1060/1070 far behind (of course it depends on second coins price). Most other algos don't need fast memory.
So right for the moment 1080 is the best choice because they cost only ~20-25% more than 1070.
