Difference was bigger for DX11 games, which were becoming a lot more common at the time. I think for anybody looking to play the latest games, the upgrade was going to be bigger than the 35% 'average' shown there, which I'm guessing used a fair amount of DX9 games in their benchmarking.
-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
AMD Polaris architecture – GCN 4.0
- Thread starter Boomstick777
- Start date
Soldato
- Joined
- 28 May 2007
- Posts
- 10,220
Yea they have a decent stack of games which is why i use them a good bit. Having a quick look through at some of the dx11 titles and indeed they do seem to gain the gtx670 a bigger lead between 40-50%.
i think you misunderstood my post, i wasnt saying polaris or vega would achieve that, my comment was about the argument of Nvidia going with big chip for flagship and AMD going for small cost effective chip, so i pointed that to me i wouldnt bother for any new flagship, unless it crosses the next wall which is 4k@60 or 1440p@120, and to do that 10-20% over Fury/Ti wont cut it, to me that wall is about 2x390 perf or something like 50-70% perf over Ti
Last edited:
To get 60fps @2160p in all games you need a 300% increase over todays flagship cards. This is not going to happen with Pascal/Polaris or possibly not even in the generation of cards that come after them.
Worst still in 2 or 3 years time when such cards are available the requirements to run AAA games will have increased by about 300% bringing us back to square one.
To get 60fps @2160p in all games you need a 300% increase over todays flagship cards. This is not going to happen with Pascal/Polaris or possibly not even in the generation of cards that come after them.
Worst still in 2 or 3 years time when such cards are available the requirements to run AAA games will have increased by about 300% bringing us back to square one.
10nm seems to be well on track so in theory 300% should be feasible in the generation after.
Wouldn't write off getting quite close on 16/14nm finfets between the extra space for transistors, architecture optimisation and extra clock speed potential a theoretical big card that was optimised purely for 4K gaming could probably exceed 200%.
Soldato
- Joined
- 25 Sep 2009
- Posts
- 10,137
- Location
- Billericay, UK
To get 60fps @2160p in all games you need a 300% increase over todays flagship cards. This is not going to happen with Pascal/Polaris or possibly not even in the generation of cards that come after them.
Worst still in 2 or 3 years time when such cards are available the requirements to run AAA games will have increased by about 300% bringing us back to square one.
Na, 1080p was a ******* to play at with decent frame rates when that started to become affordable but today you can do get away with a mid range card. The same will hold true for 4K imo.
Caporegime
- Joined
- 9 Nov 2009
- Posts
- 25,641
- Location
- Planet Earth
Not according to TPU:
https://tpucdn.com/reviews/NVIDIA/GeForce_GTX_670/images/perfrel.gif
The GTX670 was around 30% faster than a GTX570.
This is the thing - expecting a 232MM2 chip to be 50% faster than a GTX980TI is a lot.
If a 300MM2 GK104 could barely get 10% maybe 20% at most over a 565MM2 GF210,that would making it the biggest performance jump in like a decade if Polaris 10 could thrash a GTX980TI.
If you want to compare similar size dies,ie, GTX560TI which was around 330MM2 to a GK104 based GTX680 which was 300MM2 then it was 60% to 70% or thereabouts,comparing similar sized chiops.
But that is the problem here. Polaris 10 is around 232MM2. The closest GCN GPU for that is in the R9 270X and R9 370 which is 212MM2.
Let's look at the latest TPU GPU review:
https://tpucdn.com/reviews/Gigabyte/GTX_980_Ti_XtremeGaming/images/perfrel_2560_1440.png
If a 232MM2 Polaris 10 with a 256 bit memory controller was Fury level performance,it would nearly 2.25 times faster than a similar sized GCN1.0 die. If it were only R9 390X level it would be twice the performance.
So,if AMD want to get 50% over a GTX980TI,ie,double a R9 390,AMD would need to get close to 3.77 times more performance out of a similar size Polaris die when compared to GCN1.0.
Now compare that to GCN1.2,ie,the R9 380X.
The R9 380X has a 359MM2 GPU. If P10 was R9 390X level performance,it would be 50% faster and if it was Fury level performance it would be 68% faster for a 35% smaller die.
At this point if it needed to be 50% faster than a GTX980TI it would need to be 2.82 times faster than Tonga whilst having 35% less surface area.
Hence if we tried to compare GCN1.2 and P10,you would need Polaris to get 4 times more performance out of a similar sized GCN1.2 die.
At that point if AMD matches Fury with such a small chip,it is not doing that badly.
If Vega is 300MM2 to 400MM2,with HBM2 it might be the one which beats a GTX980TI convincingly on the AMD side.
Last edited:
Don't forget, the Polaris 10 is using a lot less power than Tonga, (seems to vary depending on which leak you look at) AMD themselves have said themselves (on the last conference call) that they get 2* per/W. so if it replaces the 380(Tonga) you either get twice the performance or half the power, you cannot have both and stay at 2*per/watt.
With 4 TXs it is only just possible in some games to get 60fps @2160p. Between them they pack 32 billion transistors.
What happens when you start getting 120htz 2160p panels ?
What happens when resolutions higher than 2160p start getting used for gaming ?
I don't think GPUs will be catching up in the next few years.
As to 1080p you need a lot more than a mid range card to max out ROTTR for example.
Caporegime
- Joined
- 9 Nov 2009
- Posts
- 25,641
- Location
- Planet Earth
I personally don't see even the 310MM2 to 320MM2 GP104 or a 300MM2 to 400MM2 Vega being that fast over a GTX980TI since it is a 600MM2 chip focussed towards SP performance and gaming,unlike Nvidia previous large chips. This is why it is such a monster of a gaming chip.
If Nvidia or AMD manage that with a 300MM2 to 350MM2 sized chip I would be impressed,as that would be an abnormally big jump.
Who knows it might happen,it would be nice for a change!
Last edited:
On the gaming front the top end cards will just manage to play the top end games at decent framerates, pretty much as they always have. Once the cards get faster the developers make the game look better and use more power to render, an ever repeating cycle.
Your problem is that 'maxed' out for you means 8XAA. I'm guessing that less than 1% of PC gamers bother using such a high AA setting.
The vast majority will use 2XAA at call it a day.
I personally don't see even the 310MM2 to 320MM2 GP104 or a 300MM2 to 400MM2 Vega being that fast over a GTX980TI since it is a 600MM2 chip focussed towards SP performance and gaming,unlike Nvidia previous large chips. This is why it is such a monster of a gaming chip.
If Nvidia or AMD manage that with a 300MM2 to 350MM2 sized chip I would be impressed,as that would be an abnormally big jump.
Who knows it might happen,it would be nice for a change!
TSMC 16nm FF+ seems to allow for some fairly nice performance improvements (not to mention any architecture improvements) I suspect a ~320mm2 core will perform faster than generally people seem to think.
Caporegime
- Joined
- 9 Nov 2009
- Posts
- 25,641
- Location
- Planet Earth
But 50% to 70% faster than a GTX980TI,especially consider how much they can be overclocked? Plus we have the situation the GM200 was a large 600MM2 SP core unlike the GK210,GK110,GF100 and GF110 which were designed for significant DP performance.
Unfortunately not an easy one to pin down due to the differences in architecture/implications of FP64, etc. support and/or lack of etc. but in general if you look at other vendors like Apple and Xilinx then 16nm FF+ seems to have some nice potential for actual performance increases i.e. http://www.xilinx.com/support/documentation/xcell_articles/Xcell90_p8 Cover story.pdf (obviously this is mostly per/watt and has other optimisations they've added in as well compared to their previous products).
Last edited:
I personally don't see even the 310MM2 to 320MM2 GP104 or a 300MM2 to 400MM2 Vega being that fast over a GTX980TI since it is a 600MM2 chip focussed towards SP performance and gaming,unlike Nvidia previous large chips. This is why it is such a monster of a gaming chip.
If Nvidia or AMD manage that with a 300MM2 to 350MM2 sized chip I would be impressed,as that would be an abnormally big jump.
Who knows it might happen,it would be nice for a change!
to me the chip need to be 400-500mm², for 4k to become like 1440p of today, and if we believe the rumors about nvidia's 1080GTX, will have 312mm², that would make it 20% over 980Ti, which i find pointless, what would any enthusiast benefit from an upgrade titan/Ti/fury to the next flagship ? pretty much nothing, he will still have the same issues, running 1080p at 140fps instead of 120, or running 1440p at 90 instead of 70, and running 4k at 43fps instead of 34.
so the new line up would be a battle for mainstream and market share not flagship, because the latter have no real incentive to it other than ppl who like to have the fastest card just for the sake of having it, and these are minority compared to the rest
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
The increased performance vs decreased power info on any new process is always talking about basically the maximum with best case scenarios. It's usually about chips that can operate at lower clock speeds and voltage at idle in regards to power, generally about lower power devices and isn't a generic all chips will achieve the same improvement type of statements.
Say you have a chip on 28nm that needs to run at 0.8v to run under load but can only achieve 1Ghz at that clock speed, now on 16nm it can run at 1.6Ghz at 0.8v and you can claim a 60% performance advantage, but the 28nm chip might be able to run at 2.5Ghz at 1.1v, and the 16nm chip might run at 2.6Ghz at 1.1v. So in a specific range of operation that chip sees a 60% performance advantage at a specific voltage but it doesn't across the range top to bottom. When they talk up these numbers these are the corner cases used with absolute best case numbers.
Remember that 28nm also offered 50-60% performance gains over 40nm, but did we see chips go from 1.3Ghz overclocks to 1.9Ghz overclocks from 40 to 28nm? No but we did see mobile chips gain significant clock speed at low/medium clocks at low power.
Say you have a chip on 28nm that needs to run at 0.8v to run under load but can only achieve 1Ghz at that clock speed, now on 16nm it can run at 1.6Ghz at 0.8v and you can claim a 60% performance advantage, but the 28nm chip might be able to run at 2.5Ghz at 1.1v, and the 16nm chip might run at 2.6Ghz at 1.1v. So in a specific range of operation that chip sees a 60% performance advantage at a specific voltage but it doesn't across the range top to bottom. When they talk up these numbers these are the corner cases used with absolute best case numbers.
Remember that 28nm also offered 50-60% performance gains over 40nm, but did we see chips go from 1.3Ghz overclocks to 1.9Ghz overclocks from 40 to 28nm? No but we did see mobile chips gain significant clock speed at low/medium clocks at low power.
Indeed - the results seen by various companies using TSMC 16nm FF+ gives a fairly hopeful picture for performance though - granted not many of them are working with something of the scale/demands of a GPU. (Adding onto that GP100 has a fairly healthy boost in clocks which generally for those kind of platforms is on the conservative end of what is possible).
Yup DM is quite right, these things like 2*per/watt or the new chip will beat a current top end is normally best case. Just as certain chips are better than others in different areas.
Each chip will have a sweet spot for it operating temp, voltage, speed, these companies tune these things to get the best they can and sometimes they go beyond the sweet spot just for being better in one area or another.
Each chip will have a sweet spot for it operating temp, voltage, speed, these companies tune these things to get the best they can and sometimes they go beyond the sweet spot just for being better in one area or another.