It was expected.
AMD wins in dagger-hashimoto because they have cards with wide bus - more mem controllers so faster random memory access/latency. I don't see how bigger page supported will help for "random" access
Obviously scaling on AMD is very very low after 1000-1200 shaders (7850/7870 265-270)
Its all because of bus width...
270 can make ~ 20MH/s with core:mem ratio of 2:3 (exp atm)
290 has 2x wider bus, 2x shaders... but afaik never was close to 40MH/s
Fury can make 35MH/s due to ver wide HBM memory.
So my prediction:
1070 - 20-24MH/s (the maximum for 256-bit bus)
Polaris10 - same lvl
290 did get to 30MH though, so it was still scaling somewhat - just not 100% scaling vs. shader count.
Limit is definitely somewhere in the memory system though - and it's NOT just bus width, Fury/Nano have a 4096 bit wide memory bus (due to the structure of HBM) yet they're as fast as the 290 or about the SAME hash.
Genoil IIRC was speculating it was a limit in the TLB table hardware, but I dunno how far their research into that has gotten.
I don't see anything close to 20 MH out of my 7870s - the 270 has the SAME shader count, but it does have somewhat faster memory and IIRC 2x the memory bus width, but while it's certainly a good bit faster than the 7870 it's nowhere near 2X as fast, much less 2x PLUS the memory speed ratio.
TDP target on Polaris should be a LOT lower - the smaller node is a lot more efficient, even if you puch more transistors AND kick the clock rate up quite a bit.
On the other hand, it won't be out in time to do much Eth mining with, unless AMD pushes the release dates up a good bit.
...
