Originally Posted by tajoh111
My predictions have been accurate so far on Polaris.
It's partners word on the matter and I suspect partners don't want you to buy reference cards but instead want you to buy their versions and derivatives of the rx 480. The margins are better for them.
Cards did get a frequency boost. The stock clocks of cards like tonga were 1000mhz and these usually overclocked to 1150mhz or so with a none reference cooler.Having cards now clock at 1266 and overclock to 1400 is a boost over last gen. You might want to dub it the tajoh GCN clockspeed penalty but the GCN penalty is very real as seen from GCN to GCN4 nowadays. On 28nm, GCN's clocks got worse even those the process that each GPU was made on was an more advanced and higher clocking node then the previous version of GCN. Add in Nvidia was using the same technology and process and continued to improve clocks, and it adds more evidence to my hypothesis.
It isn't just the process that determines frequency, it's the architecture as well.
1450mhz FIJI when running on ln2 should have been a redflag about the frequency potential of GCN as well as the frequencies maxwell was reaching under ln2. It real evidence as is the current clocks and overclocks of Polaris.
This guy is ranked number 4 on hwbot so he knows what he's talking about. This card was hard modded and overvolted to levels beyond safe for air cooling and even 24/7 water cooling. Aib partner cards should get really any better considering they are not going to go as balls out as this guy did.
In reviews, using the stock cooler, overclocks vary only 50mhz between the best and worst which is about 4% which means binning won't play a crazy difference. The sooner we can squash the 1.6ghz rumors, the less hype that is built up, which is better at this point.
Rumors, are what lead to the trainwreck of hype that was the rx 480. Actual evidence was already present on how polaris was going to perform but fanboys believing the most optimistic rumors were true vastly raised polaris' expectations.
The highest overclocked rx480 in reviews was the one in Guru3d and this was 1375mhz. The worst clocked ones were 1325mhz. This is only a 4 percent difference. What this means is the silicon lottery isn't going to be playing a big role.
What Dauer said was he expect most samples to overclock to 1400mhz for AIB which is safe. What he expect from better samples is 1450mhz for safe clocks. This agrees with the 4% variance as above.
We know right away your predictions are incorrect in that you claim with each revision of GCN, the clocks got slower. Except here with the newest iteration of GCN with Polaris they improved their clockspeeds by 20%. Then even you claim up to 1450MHz clocks are possible. That brings us to a total of around 38% clockspeed increase from last gen. So, not only did the process have something to do with it, they almost reached their baseline target of 40% increase in clockspeeds from switching to 14nm.
I really doubt AIB partners would flat out lie to a member of the press so he in turn flat out lies to the public. Doesn't seem to make that much sense. Sure companies like to promote their products but 1480MHz-1600MHz is a flat out lie according to you. So, obviously this wouldn't benefit the AIB partners or the press to be saying this type of stuff if it wasn't at least slightly accurate.
So regardless of your supposed clockspeed penalty, we automatically know with Polaris that this is not true. Because again, their baseline target is 40%, and @ 1450MHz you would be @ 38% from 1050MHz. While most AMD cards can do 1100-1150 we will use 1050 because its what every card is capable of. So while yes it isn't just the process and of course the architecture itself has something to do with it. We see 1425MHz possible on Air already with reference PCB.
I'm not doubting that the guy doesn't know what he is doing, but K|NGP|N tends to break all of his records running cards modded similarly, but again typically they are not reference cards that he is able to break his records with. He is using his own custom cards to do this level of overclocking. Which proves again, that even though the card was hard modded and overvolted it is still a stock PCB card. If max overclocks were able to be reached under LN2 with reference cards, why would there ever be a need for a K|NGP|N Edition, Lightning, Classified, HoF, etc. Why would K|NGP|N prefer to use his own cards?
We also see quite a variance in overclocks with the reference card (we have numbers ranging from 1330-1480MHz), which indeed indicates that binning is important and basically all cards made the cut from AMD (probably why there is a power issue with the reference cards). This means AIB cards will come with factory overclocks so the binning will be quite different.
So in the end if it indeed clocks to 1450MHz that means they increased their clockspeeds by almost 40% which was the baseline target with 14nm, and again this is with 4th Generation GCN, so there was no regression in clockspeed (as you insist and continually repeat). And of course we know AMD is going to have revisions of these cards:
Which means again, it is most likely NOT GCN that is limiting the clockspeeds entirely, it is most likely a manner of manufacturing process maturity. Which means, as the process gets more mature, they will be able to hit their full 40-50% clockspeed improvement targets.Edited by BiG StroOnZ - 7/5/16 at 6:59pm