Originally Posted by tajoh111
Polaris would have been on target to get the same decrease in frequency if it wasn't for 14nm finfet. I already said finfet was improving the clocks that could be achieved with GCN. Hence why I said GCN hitting 1266mhz and 1450mhz is a product of finfet. These clocks when we include the cooling would have never been reached on 28nm with equivalent cooling. Hence why I mentioned, why it might not necessarily be the process that's holding back AMD from reaching higher clocks, but the architecture itself.
What I was trying to get at is these clocks could have been higher if it wasn't for the frequency regression of GCN. E.g If GCN 1 was used, aka the frequency of tahiti on LN2 at 28nm, we could be seeing stock clocks be at 1500mhz with the potential to overclock to 1800mhz, much like maxwell's improvements on finfet. The problem was GCN if it was by design or by accident, doesn't clock as high with the more advanced variants. There is no question Tahiti reached the highest clocks, then hawaii, and lastly Tonga/fiji. If we look at hwbot, the highest tahiti's reached near 1800, the highest hawaii 1600 and the highest fiji reached was 1450mhz with ln2.
What 14nm is allows is those fiji ln2 clocks to be drawn out on air much like how 2200mhz may be achievable with pascal on a good enough sample. 1600mhz is beyond the clocks that were last achieved on 28nm on tonga/fiji which means I don't think we will get there even with the improvements of 14nm finfet or 16nm finfet.
Polaris got an increase because of finfet but the increase wasn't as large as Nvidia's increase. Why we are seeing clocks being limited so far and not the frequency jumps that Nvidia got are the limits of the GCN architecture.
Architecture has as much to do with frequency as the process they are made on.
At the end your actually agree with me half way. And that what I was insisting. The current frequency and limits are a product of the GCN architecture and the improvements of finfet. I never said GCN4 was going to clock worse when built on finfet, I said if GCN4 was made on 28nm, it would have shown the frequency regression pattern. Finfet counters this pattern but doesn't mean we are going to see 1.6ghz GCN chips.
Also, this 1.48ghz should not be included in the overclockability of an average sample. It is too overvolted for 24/7 usability and the author admits this and the power consumption and costs make it impractical to apply to a card. E.g AIB card or stock card + hard mod + water block = just a bit above a gtx 980 with a cost near 1070 with the water block, - 30% performance vs a gtx 1070 and over twice the power consumption.
Hence real world overclockability, which preserve the rx 480 value proposition is more along the lines of 1400mhz to 1450 on a good sample.
AMD exclusive partners have everything to gain by having people wait to buy their AIB models. AMD reference cooler has the lowest margins because it has the lowest cost and has the lowest potential for marketing. Also for partners that are not sapphire, because they are buying the cards from sapphire to rebuy, they lose part of the margin in the process. In addition because the rx 480 doesn't have too much competition in its price range, they don't have to worry too much about losing a sale to nvidia(atleast prior to the 1060 launches). Because of this, they would rather customers wait, so they can sell you a heavier margin AIB card, then to sell you a low margin reference design.
Reference cards overclocks as seen in reviews and hard modded cards by professional overclockers are much better evidence for the overclockability of polaris than some rumors and a single phrase by a person a couple weeks ago was said by the very same people clinging to these rumors, that his words couldn't be trusted and he was a shill for Nvidia.
While the architecture itself remains the same, here with 14nm it is a similar scenario with GCN 1st Generation. Smaller die chip, new process, high clockspeeds. It has a lot more in common with GCN 1st Gen cards that it does with 2nd Gen or 3rd Gen cards (where the trend for bigger and more power hungry begins to take over).
Surely it wouldn't make sense if custom cards were capable of only doing 1400-1450MHz, when there is a dry ice run on hwbot with a RX 480 and he's only hitting 1460MHz (he even says himself this is achievable with water). Obviously there is some issues with the reference card, if when cooling is not a limitation, they still are only able to achieve 1460MHz. Or in the case of the example you were using 1480MHz. I mean, if custom AIB 480's can hit 1450MHz, but a reference card can only do 1460MHz with dry ice? That doesn't seem to add up here.
Again seems more like a limitation with the reference design, rather than a direct product of GCN.
Here my point is, that here with Polaris, it is a more similar scenario as the 7970, compared to the 290X and the Fury X. The major reason why you saw a regression in clockspeeds, is because the cards became bigger with more Stream Processors, and of course as a result became more power hungry. These types of directions with GPU design will automatically as a result make it harder for a card to clock higher. You tend to see this with any big die cards. Perfect example is the stock clocks of the 980 and 970, compared to the 980 Ti. And of course the average overall overclocks changed:
Quite a variance in average Air and Water results.
So while we can agree that GCN doesn't clock as well compared to NVIDIA's architectures, the pattern you are seeing of regression in clockspeeds with GCN is more closely related to the size of the GPUs that they were releasing with each GCN revision. As you can see with NVIDIA, even NVIDIA loses clockspeeds when it comes to their big die cards and AMD's direction for quite a while has been mainly on big die cards.
Here with Polaris much smaller card, so has the ability to be more promising in terms of clockspeeds right from the get go and already we are seeing that coming to fruition. As the clocks they are achieving now with this card is far better than anything we've seen in a while from AMD.
There was another person who got a 480 to 1487MHz with an H100, so it's not limited to extreme hardmodding:
Better sample sizes could prove to be better in terms of what voltage mods can do. Also @ 1480MHz it will be faster than a 980. This 480 @ 1487MHz with no memory overclock (meaning more gains) did 67.5 fps in Valley 1.0 which is right around Fury/Nano performance. Again without a memory overclock. So it is better than a plain old 980.
I agree with you that 1450MHz is almost 40% but I do believe there is more to be discovered if not now, maybe with future AIB cards (couple months out), and definitely with revisions of the card because by then 14nm will be more mature. The only reason for this is I would like to see more numbers with AIB cards because I find it peculiar that a dry ice run with reference 480 only yielded 1460MHz but yet you seem to believe that AIB cards will be able to make it 1450MHz. Which means there is an inherent problem with the reference design. Because that doesn't seem to add up. It also doesn't make very much sense that with a volt mod 1480MHz is the max oc of the card but yet AIB cards have the ability to make it to 1450MHz stock. So with hwbot level modding, you only gain a 2% increase in clockspeeds? You have to at least agree with me that that doesn't make much sense and der8auer's results might not the end all be all.
While AIB partners have something to gain, you have to agree that it would be quite counterproductive to tell a member of the press their cards are hitting certain clockspeeds, when in fact it was just some scheme to sell more AIB cards.
Originally Posted by Majin SSJ Eric
Well to be fair, the GCN architecture has been very lackluster in clock potential since its launch (and clock speeds have steadily declined since the 7970). That said, I do agree that 14nm should allow for higher clocks than we are currently seeing but we are also limited by the fact that only reference boards have been launched and tested so far so we really can't say for sure what the absolute clock speed potential of P10 is as of yet...
The flip side of this lower clock speed potential than Nvidia's Maxwell and Pascal architectures is that generally GCN scales much better with higher clocks, so even though the ultimate speed may be lower than Nvidia, the performance for each MHz is higher...
If you read above, I have a basic explanation as to why we see a regression in clockspeeds since the 7970, it's a lot more logical than just saying, "It's GCN."
I agree that we only have reference boards right now showing clock potential, so I'm more interested in seeing where AIB cards end up because from what we see and hear right now it just isn't adding up.
I do believe that they will scale much better in the end, but at 1450MHz you are only really at Nano/Fury performance or overclocked 980 Performance. I say only, because seeing Fury X performance only 100MHz or so away is kind of disappointing.