AMD's next GPU uarch is called "Polaris"

[3DVagabond]

So, much higher than the slides imply... Why are they being so conservative then?


EDIT: Okay, I see that the slide is completely meaningless now.

The problem is the lack of reading comprehension by someone, everyone else deciding to take it's meaning as that person interpreted it instead of actually reading it themselves, and a 3rd person twisting it to somehow being dodgy and mean less than what it does.

The slide doesn't imply anything. It's there in black and white. I think once you figure it out on your own (Don't let other people twist the meaning for you) you'll see what they are claiming.

Remember their own cards are less efficient gaming than Maxwell. So, them comparing to Maxwell while gaming is the toughest test available currently.

Also keep in mind, it's one example and you can't base too much, good or bad, on one test.

 

[maddie]

So it looks like AMD is going the Nvidia route by launching bottom up with Polaris GPUs. It makes the most sense since FINFET yields are still a challenge at the foundries especially as die size goes up. The largest FINFET chip we have seen to date is Apple A9X at 147 sq mm fabbed at TSMC 16FF+. Since now its confirmed that AMD's smallest GPU is a 14 LPP GPU fabbed at GF, its very likely that AMD uses TSMC for big die flagship GPUs as they have the better yields compared to Samsung / GF.

I can see AMD's flagship GPU come in at 300 sq mm and use HBM2 and be manufactured at TSMC 16FF+. I predict a late Q3 launch. I am sure AMD will use multiple SKUs by disabling CUs for improving yields. Raja also mentioned only 2 new FINFET GPUs this year. So AMD has to make up the entire GPU stack using two chips. I speculate that the big die FINFET GPU chip will have atleast 3 SKUs (and maybe even 4) depending on yields.

This demo is a pretty small die, but it makes a lot of sense in a mobile product. Could make for great gaming laptops, even outside of the desktop market. It remains to be seen if the final die will come in at the same size, but it would be great to get a refresh in that power class as the 4 year old GCN1.0 Cape Verde and Oland GPUs are definitely long in the tooth. AMD's low end competes very well with Nvidia's if power in unconstrained, but they really don't have anything useful for someone with a PSU limite prebuilt to compete with the 750 Ti.

~100mm^2 to 300mm^2 is a pretty huge gap though. It will be interesting to see how long it takes in 2017 to fill that spot in the lineup without relying on cut down dies.

A few weeks ago, we had some speculation [myself & others] on the use of interposers to integrate a multi die GPU without the negatives of Xfire/SLI. It might even save production costs and allow a 100mm^2 and 200mm^2 die to give a very wide range of products.
100mm^2 [cut die and full die]for low end discrete and laptop together with 200mm^2 [cut die, full die, cut die X2 and full die X2]. Not really X2 as the integration is on the interposer itself.

Please usual suspects, this is pure speculation and after the insane 1st page of posts in this tread criticizing WCCFtech for posting the article in the OP [now seen as true], we don't need your "expertise and knowledge".

 

[AtenRa]

Just put this baby in a Quad Core ZEN APU with HBM 2 and say goodbye to Intel iGPUs.

Sadly we will have to wait until 2017 for that to come.

 

[3DVagabond]

I think GCN 4 is something like Maxwell for Kepler.

I expect that performance from core count may change a lot, and we could see that 2816 GCN 4.0 core GPU end up being on par with 4096 GCN 3.0 GPU.

Im wondering if this might have something with cache and delivery time of data to front end, and feeding up all of the cores, and emptying the schedule much faster than before.

I'm not disagreeing with you. I wouldn't really know. Just looking at the slide it looks like the shaders are one of the parts that are unchanged. Although they do mention "more efficient use of the shaders".

 

[Seba]

So what would this GPU be? Ellesmere? Something else?

First Arctic Islands rumored to be released (not necessarily in this order or all exactly at the same time) were supposed to be Greenland (big chip), Baffin (medium chip) and Ellesmere (small chip).

 

[MrTeal]

A few weeks ago, we had some speculation [myself & others] on the use of interposers to integrate a multi die GPU without the negatives of Xfire/SLI. It might even save production costs and allow a 100mm^2 and 200mm^2 die to give a very wide range of products.
100mm^2 [cut die and full die]for low end discrete and laptop together with 200mm^2 [cut die, full die, cut die X2 and full die X2]. Not really X2 as the integration is on the interposer itself.

Please usual suspects, this is pure speculation and after the insane 1st page of posts in this tread criticizing WCCFtech for posting the article in the OP [now seen as true], we don't need your "expertise and knowledge".

I would hardly consider the publication of the RTG data as vindication of wccftech. I could be wrong, but most people weren't mocking them for naming the next GPU Polaris, but for their non sequitur in jumping from Raja's post that Polaris is 2.5 brighter today than in 169 to Polaris will be 2.5x more power efficient than current GCN. While 2.5x might be around what you could expect from moving 28nm->14nm along with a good uarch jump, the whole conclusion didn't follow from any kind of evidence.


For your main point, that would be interesting. Nothing in the slides indicates anything remotely like a new dual GPU technology, but AMD has tried something like that in the past. The HD4xxx series had Sideport though it was never used, and it had been planned to be in Cypress before getting the axe. It would be cool if they brought something like it back, but I'm not holding my breath.

 

[Kenmitch]

So what would this GPU be? Ellesmere? Something else?

First Arctic Islands rumored to be released (not necessarily in this order or all exactly at the same time) were supposed to be Greenland (big chip), Baffin (medium chip) and Ellesmere (small chip).

Wrong thread as this one is about Polaris

Not sure if Arctic Island is still around, axed, or just renamed?

On another note it looks to be little Polaris based on comparison to the GTX 950.

 

[Techhog]

https://forum.beyond3d.com/threads/...ors-and-discussion.56719/page-22#post-1889628



http://media.gamersnexus.net/images...h/battlefront/battlefront-gpu-1080-medium.png

So the 950 should have been on pretty low load as well. This is only really bad for AMD if... Actually, that's tough to say. It basically just doesn't say much at all. There isn't nearly enough context. But you can make assumptions if you want...

 

[3DVagabond]

"Wow"?
Maybe i miss something: AMD is comparing a 16nm FinFet chip to a 12months old chip with nearly 2x times the transistors in a AMD biased game. They archived a 2,5x improvement in efficiency.

nVidia released the GTX980 11 months after the 290X, which has less transistors and archived a 1.9x to 2.0x improvement...
Would have nVidia demonstraded the GTX980 in march 2014 nobody would believed it.

Yet when AMD needs a much longer timeframe, a new process node and nearly twice the transistors, everybody is screaming "Wow".

I guess the hype train is back and now bigger and better than ever.

I think we have a stronger negative hype going right now than anyone saying "wow".

And You are being misleading calling SW:BF AMD biased. It runs well on all hardware.

 

[AtenRa]

And You are being misleading calling SW:BF AMD biased. It runs well on all hardware.

Wait to see what will happen with the first DX-12 games.

 

[Seba]

Wrong thread as this one is about Polaris

So Polaris isn't the re-badge/re-name of latest GCN? That was my understanding.

You say that Polaris replaces Arctic Islands name?

I was asking about that 30W/86W GPU if it is Ellesmere.

 

[dogen1]

it will depend on the details of the arch. If they have a GPU that maintains the compute and hardware scheduling in the older GCN then its wow. If they've been downsizing their gaming GPU like nvidia did to achieve their "efficiency" then not too impressive.

I think those power figures are total system (obviously). 86W playing battlefront at medium 60fps 1080p is pretty decent.

Seems perfect for laptops

Maxwell massively improved performance in many compute benchmarks(at least the ones that anandtech uses), so was it really a downsizing in terms of compute and scheduling?

 

[Techhog]

A few weeks ago, we had some speculation [myself & others] on the use of interposers to integrate a multi die GPU without the negatives of Xfire/SLI.

We are not going through this again, are we? We had people pulling this same crap last year, saying that Fiji would be two Tongas slapped together. It was a dumb, unsubstantiated idea back then, and nothing has changed to make it any less ridiculous. Why are some of you obsessed with this idea? How many years have people been throwing this idea around? Why is it always AMD? What is even the point?

 

[Kenmitch]

So Polaris isn't the re-badge/re-name of latest GCN? That was my understanding.

You say that Polaris replaces Arctic Islands name?

I was asking about that 30W/86W GPU if it is Ellesmere.

Now you got me confused.

Was GNC ever called something other than GCN with maybe a 1,1.2, etc. tagged on?

Polaris is AMD's next GPU technology so wouldn't that mean that there next cards would be Polaris and not Arctic Island?

 

[3DVagabond]

Nope..I am pretty sure pascal/GTX970/980 successor will beat TITANX around 40%.
If AMD want be competetive they need GPu around 300-350mm2 on 14/16nm

That's a guess. Has nVidia said anything?

 

[Seba]

Was GNC ever called something other than GCN with maybe a 1,1.2, etc. tagged on?

No. But the GPUs themselves had "islands" names.

Like Fiji XT (with commercial name Radeon R9 Fury X or Radeon R9 Nano) was based on GCN "1.2".

Similarly, Arctic Islands will be based on a micro-architecture called Polaris (and not GCN 1.3 or GCN 2.0).

Unless the Polaris name replaces Arctic Islands name and not GCN name.

 

[MrTeal]

Now you got me confused.

Was GNC ever called something other than GCN with maybe a 1,1.2, etc. tagged on?

Polaris is AMD's next GPU technology so wouldn't that mean that there next cards would be Polaris and not Arctic Island?

Officially no, 1.1, 1.2 were just things added by the press to separate different GCN feature sets. Considering RTG is now calling this 4th Gen GCN, Tahiti would have been GCN1, Hawaii GCN2, and Fiji GCN3.

From the second page of the AT article, I get the impression that the whole stack of dies based around GCN4 and the other common aspects of GPU will be called Polaris. Individual dies might still have names based (possibly) on Arctic Islands.

 

[3DVagabond]

http://www.hardware.fr/news/14456/amd-annonce-polaris-future-architecture-gpu.html

Il est équipé de mémoire GDDR5. Nous n'avons pas encore de nom pour ce GPU et n'avons pu l'apercevoir que brièvement sans pouvoir en prendre une photo. Tout juste de quoi apercevoir qu'il s'agit effectivement d'une petite puce et d'un packaging compact.

Thanks for posting that. I really like HW.FR's coverage of things. I wish they had maintained their English site, but Google.trans will have to do.

 

[3DVagabond]

This image has been going around. Do you guys remember the name of the technique that Nvidia uses to save power that is wasted processing frames that are not needed? I assume that when a game is VSYNC capped at 60 fps, it is saving power vs running at 150fps or whatever max the hardware is capable of hitting. Are there any benchmarks that compare 60fps capped power consumption vs uncapped on a 950? If we had that info we would be able to better assess just how good Polaris is.

Yeah. They give us the info. It's obvious from the overall power usage that the rest of the system is coasting along or else it would use more power than that not counting the GPU.

 

[3DVagabond]

Did any of you actually read the Anandtech article???

They specifically called out that it was capped at 60 fps. AMD also similarly in the last 6 months released a new driver with frame rate capping capabilities, and boasted about how it increased perf/watt. They picked a 950, not a 960, so that more of the chip would have to be spun up to higher frequencies to hit 60 (wider w/ lower clock = more power efficient, narrower with higher clock = less power efficient). The new chip can likely do more than 60fps in SWBF, and the 950 is probably nearing its cap, and this was intentionally chosen. Do the math.

Those numbers they showed are certainly true, but they've also picked the precisely best light to show it in. As anyone should expect from a products company marketing its product...

Do we know this is what they did, or is it just a possibility showing the worst case scenario?

 

[3DVagabond]

Look at this Image , both are capped at 60Fps.

http://www.overclock.net/t/1586552/...t-polaris-gpu-architecture/0_40#post_24760011

850Mhz at 0.8375 volt.

Now if we look at Slide :



totally different.

Well, there's no way the nVidia card and the AMD card could use the same drivers. I'm sorry, but I'm missing your point. Could you explain, please?

 

[AtenRa]

Do we know this is what they did, or is it just a possibility showing the worst case scenario?

Also, SW Battlefront is currently DX-11. Polaris in DX-12 games will have even higher perf/watt vs 28nm GCN and Maxwell.

 

[AtenRa]

Well, there's no way the nVidia card and the AMD card could use the same drivers. I'm sorry, but I'm missing your point. Could you explain, please?

Yea i believe AMD used one driver older than what its available today. But i dont see anything new for SW Battlefront in the latest driver that could make things better for the GTX950.

The driver AMD used was released on December 1st 2015 and the latest one was released on 21th of December 2015.

 

[3DVagabond]

at bit OT but I find interesting to mention that all Polaris official slides are branded RTG (Radeon Tech Group) without any mention of AMD. First time it happens. A sign that RTG will become an independent company (aka ATI) ?

It could just be a separate division. You don't see GMC mentioned on Chevy truck adds.

 

[3DVagabond]

Would love to see RTG spun off as a separate entity.

Depends on how Zen does and also don't forget the consoles. Those are actually under the CPU division. That could really screw things up if the "two companies" had mutually exclusive goals.