I have the sapphire reference. I modded the bios with polaris bios editor-> core clock, mem clock and mem timings (the straps). Most of people use the 1500 mhz strap and copy it to higher mem clock straps. I found that for my card 1375 memory strap is best. it's stable at mclock 2010 => 780h/s.
with nitro cards you should be able to go higher , the mem is better( although the latest nitro 4GB comes with elpida memory instead of samsung and is not supported yet by polaris editor).
for undervolting you have to use TRixx or wattman or other. I run at core voltage 825 mV and mem controler 981mV. you should go way ver 800H/s. I'm looking forward to see your results.

more info here: https://bitcointalk.org/index.php?topic=1574652.0

Edit: just noticed you have Rx 480 then better see there: https://bitcointalk.org/index.php?topic=1584617.0

I did a lot of tests to figure out witch voltage is wich and got only confusing results. There is clearly some synchronisation, when you change one voltage the others are often modified too.

Does someone have a clear idea how the whole thing works?

Maybe check temperature/cooling, the clock goes down when you reach the limit.

Tried a different approach: lower mem clock but with strap from 1375 mem clock. Reached 800h/s cc 1200 / mc 2025.
Not stable yet.

It was stable over night, still running, but i'm about to try over settings.

little correction, this power consumption is for cc 945mV/ cm 975mV. After some modification the reload of the setting with trixx modified it a bit and I did not notice.

GPU-Z shows ~85W and I have 185W at the wall for the entire system (iddles at 80). As a comparison, mining ETH it uses 205-210W.

You can find some in some threads in this forum (sorry don't have kept link).
But the one I use now, I made it by myself.

All you need is polaris bios editor and ATI Flash.

But carefull my first flash resulted in a bricked card. Fortunately had an old quattro in second PCIe to use to unbrick.

Hi,
I tested on RX 470, there's 25-30w less on the wall while mining XMR. Whole system a lot more stable.
The obvious difference I can see in GPU-Z is the use of the MEM controler 100% (eth with claymore) vs 25-30% (XMR with claymore).

Hi there,
my sapphire rx 470 OC 4gb (ref) mines stable at 759h/s with modded bios (1500 timings and more) cc 1150 / 950mV - mc 2090 / 981mV
did not play with -h for now.

Hi there,

I'm the happy owner of a sapphire 4gb ref with ASIC quality 90% (somme say it's important...Duno).
I can do a lot overclocking/underclocking/undervolting. Never tester overvolting, I try to spare electricity.

I tried A LOT of settings playing with Trixx (most stable) and WattTool (most practical) mostly running claymore's. Often both together (tricky but...)! Homemade custom bios (I may have flashed a undred bios last week, my wife and kids where abroad ;-)

Usually I dont go under 981mV for memory controler, it's not worth it vs stability.

For GPU voltage I should stop at 925mv under that mark hashrate drops or/and unstable.

For memory clock, hashrate drops over 2090/95. If somebody has a good advice about that i'll be happy to try.

This said, I reached up to 29,9Mh/s dual mining decred (dcri 10) and very briefly once only 30.02Mh/s but I was not happy with the stability and the share count did not really improve. Droped it.

Most stable in the time is more around 28,8 with clock at 1130 and memory between 2025/2050.

Some higher marks seem stable, but over several ours the share count drops by 5 shares every hour until it becomes ridiculous. I'm still looking for the winner!

To finish, GPU-z shows between 85 and 100 W depending of the settings, but at the wall add at least 20W more and probably 30W according to my powermeter.

It just crashed while writing at 1150/908mV-2050/981mV-202W whole system. It run for an hour , then dev fee provoqued a deconnection. At reconnection...crack.

I cant get it running stable more than a few hours under 210W/rig at the wall. Sometimes it's so on the edge of stability that I wonder how it can continue computing...

Ps: I intended to run it on linux at first, but 40W in excess vs windows stop this experiment!