Here's my prediction: - Tonga will be 470, Hawaii will be 480, Fiji will be 490. 460 will be little Polaris and Rage will be big Polaris However, I have a feeling that Hawaii might be replaced or removed from the lineup.
The last time that they switched (when GCN was introduced), they kept selling the older cards, but as they were.
"2 new GPUs in 2016" is coming straight from AMD, if my memory serves me correctly. So it's this small Polaris for laptops and the other one is proly a little-big die ala Tahiti/Tonga for high-end. With Tonga in the middle and Bonaire at the bottom. As for Nvidia several GPU codenames had already been spotted in their driver: http://www.3dcenter.org/news/reihen...cht-gp100-gp102-gp104-gp106-gp107-gp10b-gv100
I personally think there'll be a big chip like the FuryX, and a 'cut down' one, like the Fury. the rest will be rebrands. Would fit with '2 new GPUs' and stuff.
There are no rebrands. Why do people keep spreading these lies? All FinFet Chips will be Polaris architecture. No Tonga crap anymore.
rebrand is just to use the same gpu, if you tweak the gpu if not a rebrand just and improved version, if AMD brings all the current gen to the new process it will be a improved version more powerfull and optimized.
http://www.guru3d.com/articles_page..._january_2016_amd_polaris_architecture,1.html That's 3 new chips. Why bother shrinking Tonga or Fiji to a smaller node when the engineering effort required is the same as using a new architecture? The entire Rx 400 series is a new product slate. If they wanna keep selling Tonga/Hawaii chips, then only with the Rx 300 series label.
thats from wcftchwctf (Aug 2015) this is more recent and straight from their main GPU man (Raja Koduri): Raja also talked about how Advanced Micro Devices’ RTG will need to execute on their architectural designs and create brand new GPUs, something that Advanced Micro Devices has struggled with lately. He promised two brand new GPUs in 2016, which are hopefully going to both be 14nm/16nm FinFET from GlobalFoundries or TSMC and will help make Advanced Micro Devices more power and die size competitive. Having more products to sell is a good thing for Advanced Micro Devices because there’s no getting around that some of AMD’s GPU architectures have gotten quite long in the tooth. http://www.forbes.com/sites/patrick...ies-group-raja-koduri-talks-about-the-future/
If you take Fiji and die shrink it to 14nm and replace GCN 3 cores with Polaris then that is a completely new GPU and cannot be called a rebrand. If Raja's statement is correct then there will only be 2 14 nm GPUs this year and we have already seen one of them, the entry level one. To compete with Pascal, there has to be a high end flagship GPU this year (Greenland?). May be the rest of the lineup will follow next year and we may not have a single rebrand (read 28 nm) in the 400 series. Interesting times ahead.
Hardwareluxx spewed some extra info These low power ones are indeed lower end gpus with GDDR5x or GDDR5 coupled with 4790 @ win10 https://translate.google.com/transl...d-demonstriert-polaris-system.html&edit-text= But this is just a lower end, HBM2.0 will be reserved for high-end just like by NV Pascal.
Thanks for the link. Now the Polaris ad test make more sense. What i fear the most is the part saying only 25% of the power saving are active in driver and firmware. From what we saw with Crimson agressive power saving difficult to disable even with 3rd part software (ClockBlocker) this is not something working well in the performance side. Imagine x4 power saving features in drivers without a proper way to disable them. Power saving dream could also be a true performance nightmare.
First of all, let's call it mildly improved versions of old GPUs without a new architecture, if you fancy that more. And to be honest, I wouldn't bet money on all the 4xx cards having 14/16nm GPUs. Doesn't have to be bad either, since the 'old' and 'not top notch cards' would still be good and dirt cheap for people minding a budget.
While Raja Koduri may have mentioned 2 GPUs, that in reality may mean 4 Cards as we can expect cut down versions. With 3rd GPU you have 6 cards to cover entire spectrum, easy. But thing is: - AMD stops manufacturing certain 28nm GPUs which they intend to replace with 14/16nm. Let's say low-end to average GPUs where are big sales. (as we saw this GPU "benchmark") - > They release this GPU mid year as older generation stock runs low to minimize losses. - > This low-end to Average GPU will eat 45~65W (Cut~Uncut) and perform as current 90~140W GPUs. And as results AMD reenters mobile market properly. - At Q3/Q4 AMD releases GPUs with HBM2 as early release would not be economical due to HBM2 availability - > HBM 2 GPUs will have around Fiji transistor count. Eat like 150~200W, but will clock bit higher and HBM2 bandwidth will grant 10% performance premium clock to clock vs Fiji and in total being 20% stronger than Fiji - This leaves current market for R9-380(x)/R9-390(x) practically untouched as they will deliver good performance for price, and only downside will be higher power consumption - AMD replaces R9-380(x)/R9-390(x) Q1 2017 with one Cut+Uncut chip - And then Q2 2017 releases 12~14B chip as bigger upgrade over Fiji replacement from Q3/Q4 2016, delivering additional 30% of performance in 250W TDP.
There is no "simple die shrink". You have to do double the work to take a chip from 28nm to 14nm. Their 14nm chip is Polaris and the Polaris family in general. There will be no 14nm "rebrands", it makes no RnD sense at all. That's a good point. If the clock gating is not fast enough (because if it is, you don't really care), then it could be problematic. As for the whole transition, I believe they will do exactly what they did when the 7970/7950 came out. No "rebrands", just a new line of cards along with the old ones. Don't forget that the 14nm middle range parts will probably be much cheaper to produce if the yields are ok.
In regards of Polaris maybe relevant: Found in GloFo Webinar: Challenges and Solutions for 14nm Digital Design Reference Flows
This image is crazy. While it somewhat confirms that shown polaris chip is not that much under-volted. It shows extremely small leakage. Those "Relative" numbers are bogus, because person who "calculated" them knows nothing about ratios. Normalized: 28SLP: 100mW draw on chip makes 44.3mW leakage 28HPP: 100mW draw on chip makes 56.7mW leakage 14LPP: 100mW draw on chip makes 6mW leakage This kind of small leakage is crazy. And if true, then 10nm can deliver only decreased need for voltage which means it will not bring as big improvement as 14/16nm. Unless voltage required drops significantly. And those clocks... batsh*t crazy. But I really wonder why intel is clocking their IGP in x5-z8500 Atom to 600MHz Max. (Well, I know that once fully loaded this iGPU eats around 6 Watts. So those 2.4GHz would spike its power consumption close to 96W if we used exponential scaling for clock effect on power consumption. If linear, then only 24W, so it will be somewhere in between.) Edit: And I think I get what they mean by those show cases: 28SLP: eats 158mW @ 0.97GHz per certain transistor count - 163mW @ 1GHz 28HPP: eats 210mW @ 1.17GHz per certain transistor count - 179mW @ 1GHz 14LPP: eats 310mW @ 2.41GHz per certain transistor count - 128mW @ 1GHz (But in reality that last scaling may be as low as 22mW @ 1GHz.) And if we jump to total performance/Watt: 28SLP: eats 158mW @ 0.97GHz - @250W - performance coefficient of 1 28HPP: eats 210mW @ 1.17GHz - @250W - performance coefficient of 0.91 14LPP: eats 310mW @ 2.41GHz - @250W - performance coefficient of 1.27 (because while it allows for 2.48x higher clock, it eats 1.96x more power at that clocks) !!! but if you extract performance through doubling frequency, you can cut transistor count to 1/2 !!! So in the end those 14nm chips may not be as expensive as I would have guessed till now. And while that coefficient of 1.27 of real world performance/watt improvement looks small, it is at peak clock. I guess somewhere around 1.5GHz and beefy chip, it may deliver 1.6x more performance than today's 250W cards. Edit2: Bad thing - Thermal Density That showcased set of transistors occupies 3.7x less space (surface area of chip), but draws 1.96x more power. That means up to 7.25x higher Thermal Density. I can see big chips under water. I apparently do not mind, but not everyone will like it.
All these stand with specific transistor configurations. That's why I believe that the data they gave vs the 950 is so important.
