Inspectah_Deck
Member
Hmm, I´m no expert, but thats exactly what Global Foundries proclaims about their new 14 nm process:
http://www.neogaf.com/forum/showpost.php?p=203217421&postcount=1045
AMD Polaris architecture to succeed Graphics Core Next
- Thread starter wachie
- Start date
http://www.neogaf.com/forum/showpost.php?p=203217421&postcount=1045
Inspectah_Deck
Member
Polaris is for the <300$ market.
Aztechnology
Member
Because not everyone wants to spend $600 on a GPU?
http://www.tsmc.com/english/dedicatedFoundry/technology/16nm.htm
This was likely only part of it, but it was probably a significant part, considering the fact that higher clocks and lower power consumption seem to be the only real differences between Maxwell and Pascal. And yeah, I do realize that this may not be referring strictly to clock speed. However, I know that the reason is FinFET's faster switching, which does mean higher clocks.
Check the TPU review I linked, 980Ti is only 10% faster than Fury X at 1080p.
2560sps @ 1.6 GHz matches Fury X Tflops. A small increase in utilisation gives you > Fury X performance. A more realistic 1.4-1.5Ghz target with the utilisation gives you Fury X performance.
R9 270 is 150 Watts, multiply by 2.5 and you get > Fury X performance.
P10 is rumoured to be around 230mm^2 so a little bigger than Pitcairn. Perf / mm is about 2.2x higher so again multiply by R9 270 by 2.2 and factor in the slightly bigger die and you are at > Fury X performance.
The 7870 was 20% faster than the 6970, P10 is a lot like Pitcairn and Hawaii is a lot like the 6970, a 20% increase over Hawaii would be in the Fury X ballpark.
When all roads point in one direction it tends to be the right area so expecting near Fury X performance from P10 is not far fetched.
Phawx
Member
AMD needs to announce VEGA10. Polaris mid-range is great for the market, but people will see it as another 390 just a bit cheaper.
It doesn't matter if only 10% of the market buys GPUs > $500, 100% see who holds the crown and then determine a winner.
HBM2 in Oct would be expensive, but they need to make the move. They need another 5870.
It doesn't matter if only 10% of the market buys GPUs > $500, 100% see who holds the crown and then determine a winner.
HBM2 in Oct would be expensive, but they need to make the move. They need another 5870.
frontieruk
Member
So a card punching in at a 980ti spec for ~$300 isn't going to sell? What if said card is outperforming the 980ti in most of the game benchmarks of recently launched games? People look at the top but they also look at the prices that they are willing to drop money for, if the amd is faster than the opposition in that range things could swing.
Phawx
Member
Seeing that GPU is 80-20 in Nvidia's favor, I have a feeling many people are going to be looking for cheap 970s over Polaris.
opticalmace
Member
980 Ti performance for $300 would be pretty nuts. I would upgrade from my 970, that's for sure.
I don't really think it's going to happen.
SneakyStephan
Banned
Yeah the midrange market is going to get absolutely flooded with cheap second hand gtx 970s (mine probably one of them, depending on how 1070 benchmarks turn out and the price for an aftermarket cooler one in my country)
Amd are going to have to compete with those because a 150 dollar second hand 970 is pretty appealing if you didn't have a decent gpu yet. (they're already 220 dollars new I saw in the other thread) and probably a lot less than what amd is going to charge for a 390/970 equivalent polaris 10 card
Phawx
Member
From the conversation I had with the AMD GPU guy a PAX East, it's possible. From what the dude said to me, they aren't even using GDDR5x on Polaris. It's supposed to be very low power, low price a bit better than a 390.
Now how low price is the question. I think the third party manufacturers would be a bit upset that their entire stock of 3xx series will have to basically go on firesale or Polaris will be $350ish.
This is pretty much why AMD needs to announce VEGA10 in Oct. Polaris is going to be good from a lot of angles but as much as AMD might want to cannibalize the 3xx series, how are the partners going to feel about that?
dgrdsv
Member
Process characteristics by themselves don't mean much, you have to be able to take advantage of them. With AMD saying a lot about power efficiency I'm pretty sure that their clocking will be conservative - GCN isn't very good at handling high clocks and they'd need to do a big rebuild in vain of Kepler->Maxwell to get to where Pascal is. There's a possibility that Polaris will do it of course but for now there is no signs of this. The early Polaris samples were running on a very low frequencies - 800-900MHz - and I'm not expecting more than 1200MHz from the final silicon. But as I've said would love to be wrong because otherwise we're not going to see a 1080Ti this year.
You can't multiply the performance by power efficiency gains as that works in the opposite direction: a new chip will provide the same performance as the old one did with 2x less power consumption. When you start ramping up the clocks and unit numbers you're loosing power efficiency much faster than you gain performance.
Hari Seldon
Member
So the Polaris and 1070 might be neck and neck in terms of both performance and price?
This particular characteristic only applies within a process node. Each node (and architecture) has different base frequencies where they operate at optimum efficiency.
dgrdsv
Member
If you'd look at what I was replying to you would see that this is exactly what Marlenus is doing - he's taking a 14nm chip and expanding it into a bigger one translating the suggested Polaris power efficiency gain into a linear performance gain. Which isn't how stuff works.
Hesemonni
Banned
Duh, yeah
My bad.
Phawx
Member
Potentially.
I not working from the efficiency gains. I am using AMDs performance / watt metric that states for a given wattage performance is increased by 2.5x. Now that is probably a best case scenario so I doubt Polaris is 2.5x the perf/watt of Fury Nano.
As far as changing clocks the 2.5x metric will include things like frequency changes, architectural improvements, design improvements, leakage reduction techniques etc. Unless 150 Watts is too high for Polaris 10, or the 2.5x perf/watt improvement is an extreme edge case then there is nothing wrong with using it to predict performance.
Then there is the fact that the other scaling methods, like perf/mm^2, and previous node shrinks also converge in the same performance level bracket and you can increase the confidence of the results provided the starting conditions are close to accurate. As those are rumours this is more a best guess than a concrete estimate as those rumours may be entirely false but until further information comes along it is the best info we have to go on.
As far as I see it unless Polaris is smaller or lower power than the current rumours indicate the ballpark of Fury X performance is entirely possible.
The 1080 is around 80%+ faster than the 980 which was already a massive perf/watt gain over the 780Ti. AMDs cards, excluding Fury, are in the same perf/watt envelope as Kepler and AMD has not made any power saving improvements at an architectural level so has catching up to do. Judging by NVs gains, AMDs claim of 2.5x perf/watt seems entirely doable as a combination of node shrink, process change and architecture improvement so I see no reason to think this figure is an extreme edge case probably the general case at a given power envelope.
hooijdonk17
Member
Basically because there are people like me who want a great ~100$ GPU that can give 60fps 1080p for current gen games and it will sell bucketloads. That's for Polaris 11 anyway. As for Polaris 10, because it will probably destroy all competition in the 200-300$ market..
Hi, yes, this is me. I don't have much spare money in life atm but want to kick my 660 up the arse.
Aztechnology
Member
This, this right here is the reason why I created the thread about Adaptive Sync and why Nvidia has let me down so much with it.
SneakyStephan
Banned
That whole thread (not your OP) is a big mark of shame on the neogaf pc community, just embarassing how vehemently people will argue against their own interests and those of their fellow gamers once they have invested some money into something.
hooijdonk17
Member
Ok now I need a link..
Aztechnology
Member
Beat me to it. Yea that's the one.
Yeah, one thing I have noted about Nvid vs AMD theorising is that whoever has THE most powerful card of a generation will be the winner. Nobody seems to consider the budget enthusiast, or the mini-ITX gamer.
E-Cat
Member
Why drive anything but a Porsche?
Aztechnology
Member
I appreciate that, you had a really good post in there. I really cannot for the life of me understand why as a community we shouldn't be pushing for more options. I did not anticipate the reaction I got from so many people there.
Because then you can create an AMA on GAF and get everyone salty about how wealthy you are.
hooijdonk17
Member
Holy shit, that thread.....
dgrdsv
Member
You do. And I'm starting to think that you can't write as well as you can't read.
You can't use power efficiency gains to predict performance, it doesn't work this way. You are likely to get the same performance for 2-2.5x less power but this is most likely something which will be true only for P11 as the lowest and coolest chip in the lineup.
Yes, there is a lot wrong with using this to predict performance as if you're thinking that you'll get 2.5x Fiji performance from a Polaris GPU of Fiji's size - you will be severely disappointed. The power gain efficiency goes down much faster then you gain performance with bigger, more power hungry GPUs. We're lucky if we'll get even 1.5x for a top end piece.
But I think that at this point it's better to just wait a month and we'll see where Polaris is from benchmarks.
E-Cat
Member
I fully expect a > 1.5x boost over Fiji on 14/16nm. It may not come until next year, but anyway-
I will use the HD 7850 vs Fury X as an example. The HD 7850, at 212 mm^2 and 130 W, is probably pretty close to what the P10 chip is going to be.
Radeon HD 7850 - 212 mm^2, 130 W, 1.76 TFLOPS
Radeon R9 Fury X - 595 mm^2, 275 W, 8.60 TFLOPS
Mind you, we're on the same 28nm node. And yet the cards have a 4.88x difference in raw TFLOPS. Due to architectural improvements, it's even more than the linearly extrapolated die size factor of 2.81x or the 2.12x difference in TDP.
So, let's say that the 232 mm^2 Polaris 10 GPU is 6 TFLOPS. If AMD releases a 600 mm^2 Polaris card, do you really expect it to be only 12-13 TFLOPS?
AMD have not said you can get X performance for Y% less power, an efficiency metric. They have said for X power you get 2.5x the performance, a performance metric. Probably based on low or mid range cards like P11 vs Carpe Verde/Bonaire or P10 vs Pitcairn. This is an increase from their previous 2x perf/watt claim so what prompted that? Real data perhaps from their first batch of samples?
Do you really think a hypothetical 600mm P10/Vega will only be 50% faster than Fury X. That would be shocking as the 330mm 1080 is basically already there and would be with an overclock. AMD are nowhere near that far behind, if they even are behind.
Rather than criticising other peoples reading skills perhaps you should learn how to do it.
E-Cat. If the 2560 SP is correct for P10 6 Tflops would need a 1.2Ghz clockspeed. The clockspeed seems a bit low but it would be a bit above the 390X from a pure compute POV.
E-Cat
Member
Yeah, I was thinking 2304 SP @ 1.3 GHz. I guess we'll know soon.
dgrdsv
Member
Rumor: Polaris validation Failed Might Launch in October Now
Well I certainly hope that this is bogus.
If only flops would convert straight to performance. But they don't as you can see from 390X sometimes hitting Fury's performance level.
Another issue is that direct comparisons between Polaris and big Vega (since there won't be a big Polaris) may not be possible due to a memory subsystem differences.
Yeah, I think we should just wait at this point. Going off what AMD said is pretty pointless as we don't know to which any of this really applies.
Yeah, they really haven't.
Well I certainly hope that this is bogus.
If only flops would convert straight to performance. But they don't as you can see from 390X sometimes hitting Fury's performance level.
Another issue is that direct comparisons between Polaris and big Vega (since there won't be a big Polaris) may not be possible due to a memory subsystem differences.
Yeah, I think we should just wait at this point. Going off what AMD said is pretty pointless as we don't know to which any of this really applies.
Yeah, they really haven't.
Papacheeks
Banned
So are they going to release a 490 next month with the 480,470?
Or is the 490 for October?
Or is the 490 for October?
Tacklemacain
Member
No, Polaris 10 and 11 are the 480(x) and 470(x). 490 should release in October or January with HBM2.
Small Vega is probably going to be used for the 490s. Only thing we know about that so far is that it's due to late 2016/early 2017 and that it'll consist of two chips like Polaris. Some rumors showed up saying (small) Vega was pushed forward to October, but we all know how rumors go.
Jessie.Bristol
Member
I am the same way
. I am all set to possibly buy an Asus GTX 950 (without a 6-pin connector) for $135 but waiting on what AMD announces this month with regards to Polaris 11. I personally feel that if AMD really wants to capture the mainstream then they need the 460 to compete with the 950 at the 750 ti price point. And they need the 470 to compete with 960 at the 950 price point. That will allow them to lock up that segment of the market since Nvidia gave them the opening by not focusing on it (with 1060 and 1050 gpu's) until possibly the Fall. But will they, probably not, but I hope. I am currently rocking a GTX 750 (not Ti) and would like something a bit more gaming friendly.
Papacheeks
Banned
So do we know where the 480 will align performance wise? I read an article stating that the 480 would be similar to 390x performance but for 300$ or less.
This roadmap looks to have a pretty solid perf/watt claim. Granted we don't know how they got that figure but it is a specific claim which we can use to estimate the performance window of 14nm GCN parts.
I'm gonna need to see a link on that
Nobody knows.
AMD did allude to the VR pricepoint being out of reach ($330+), so going by baseline VR performance we are guessing 390-390X for $250 or lower.
Tacklemacain
Member
I think polaris 10 will be between the 390x and the fury x. If not their would be a huge performance gap between polaris and vega.
Just plain sad.
frontieruk
Member
AMD said:
A few months later the press started reporting Amd had upped their expectations to 2.5x
dgrdsv
Member
Well, yeah, perf/watt means that you get the same perf for different watt and that's what this roadmap is showing as presumably the only reason Vega is above Polaris there is because of HBM being more energy efficient.
Let's wait for Polaris reveal at least. What I expect from it is Hawaii's performance for 2x less power consumption - and that would be 2x perf/watt.
That would actually be perfect for my build to use as a stepping stone to the HBM2 cards. I'm looking at a 1080p/144hz freesync monitor, which would also be a good complement to the card.
Bullshit. Nothing more to say.