AMD: Polaris 10 targets mainstream desktop & high-end notebook

[Erenhardt]

There are some situation where Polaris is incredibly fast. Faster than anything in the market. The secret is probably that special culling mechanism in the hardware, which helps the GPU to effectively cull those false positive primitives that aren't visible in the screen. Today's hardwares can't do this.
Single wavefront perfomance is also incredibly good. 10-100 times faster than anything in the market. This is good for VR.

You did not just blow a whistle for this cho-cho hypetrain... damn, you did.

 

[R0H1T]

this is what im saying it would be stupid not to have a card that can perform up till fury lines (a bit lower just not to kill the fury cards D: () )
also why the 11? i thought the 10 is the big one and the 11(the first live demo) was the low/mid one

Maybe it's because it is in the PS4K or NEO

 

[maddie]

You did not just blow a whistle for this cho-cho hypetrain... damn, you did.

That is way beyond what anyone here speculated. Zlatan jumped on the bullet train.

 

[antihelten]

vega will surpass fury x and what is gonna fill that gap between 390x and fury?

For what it's worth AMD doesn't really need a GPU to fill that gap, since it's a really tiny gap (fury is roughly 10% faster than 390x).

Of course this doesn't mean that there won't be a GPU that falls in that spot, but if so it would likely be as a result of a larger gap in the Polaris/Vega lineup, and not the current gap between 390x and fury.

 

[airfathaaaaa]

That is way beyond what anyone here speculated. Zlatan jumped on the bullet train.

we know that amd has a hardware culling on this gcn gen they already told us lol(remember the first slides of the gcn arch for the polaris saying about the new geometry cpu? yeah this one..)

 

[Arachnotronic]

A 390X class product with much lower power consumption would be good for desktops and more importantly for gaming laptops. AMD needs to pick and choose its battles and it would seem that Polaris 11/10 target sensible design points.

 

[zlatan]

this is what im saying it would be stupid not to have a card that can perform up till fury lines (a bit lower just not to kill the fury cards D: () )
also why the 11? i thought the 10 is the big one and the 11(the first live demo) was the low/mid one

Typo sorry.

 

[zlatan]

I don't tell anything about PS4. But If there will be a PS4 "Matrix or something", it won't use Polaris. It should be binary compatible with the existing console.
Polaris uses a fully new coding scheme (ASIC 130), and an earlier GCN binary is not executable on this microarch.

 

[Adored]

Zlatan have you heard anything about the new PS4, or future console, having some kind of "dual GPU" setup? Just a theory I have brewing.

 

[maddie]

I don't tell anything about PS4. But If there will be a PS4 "Matrix or something", it won't use Polaris. It should be binary compatible with the existing console.
Polaris uses a fully new coding scheme (ASIC 130), and an earlier GCN binary is not executable on this microarch.

Do you see the PC drivers for Polaris as being fairly well optimized?

 

[maddie]

I guess the FPS min-max variation in Polaris will be much less than present cards.

 

[Slaughterem]

Lots of options.

Vega 10 SKUs (7950-7970-7970Ghz)
Or Vega 11.

A lot of people are trying to position Pitcairn 7850-7870 successor as a competitor to GP104. Even if Polaris 10 matches Fury X, it's not fast enough for 1080. That's why they aren't really direct competitors and I expect the pricing to reflect this.

So lots of speculation going on. It appears that GP104 will be 294 mm2 from the leaked pictures. 15.35 X 19.18 is the guess. If that is true then on the 16nm FF with 300 mm wafers and if it matches the reported defect density of 0.2/sq cm that GloFo at 14nm is claiming http://www.anandtech.com/show/10272/samsung-foundry-updates-7nm-euv-10lpp-and-14lpc
they will have 184 die candidates but only 57% yield for a total of 105 good dies.
GP104 is rumored to have 2560 cuda cores for the 1080 and in order to improve die yield will have 2304 for 1070 and since there is a rumored 3rd card probably 2048 for a 1060.
Another rumor indicated that GP104 will be 30% faster than each of the previous cards.
If they are able to harvest 150 dies with this strategy and each wafer costs $7500 then the 1080 chip will be $71 with the other chips being about $50.

Polaris 10 is rumored to be 232 mm2. We do not have any photos of the die yet but if 232 is accurate then we can guess that the size will be 13.4 X 17.3.

On a 300mm wafer that equals with the same defect density they will have 239 candidates at 64% yield or 153 good dies.
Polaris 10 is also rumored to have 2560 SP along with 2304 but we have not heard of any rumors of 2048.
So lets use the same idea that they will harvest 175 dies for the 2. The high end chip will cost $49 and the other about $42.
I don't think that 232 mm2 will be the size of Polaris 10. I think that it will be 265 which on 14nm equals 294 16nm. This would put the cost at $60 for 125 good dies and $50 for 25 cut dies.

From this point forward the games begin as to performance vs power. I think Polaris has the advantage due to GloFO 14 nm process being less costly per die and AMD is going to go the route of better performance per watt at a lower price but could rival the performance of gp104 if they choose to do so.

Nvidia with the rumors recently about prices are trying to goat people into buying what would be the best performing game cards but at higher frequencies and higher power. We will know more in just a few weeks, of course the fan boys on each side will give me plenty of pop corn time and laughs. Looking forward to this new round of tech.

 

[USER8000]

Supposedly the GP104 is 350MM2. Even if you were to give a 10% density advantage to AMD,that makes the GP104 close to 40% bigger.

Its my opinion ,OFC,but I think AMD even getting close to the GTX1070 would be an achievement if Polaris 10 is 232MM2.

It might explain the rumoured pricing for the GTX1070 and GTX1080.

 

[AtenRa]

I don't tell anything about PS4. But If there will be a PS4 "Matrix or something", it won't use Polaris. It should be binary compatible with the existing console.
Polaris uses a fully new coding scheme (ASIC 130), and an earlier GCN binary is not executable on this microarch.

Are you saying it will be worse than going from Hawaii to Tonga to Fiji in DX-12 game optimization ???

 

[S.H.O.D.A.N.]

Are you saying it will be worse than going from Hawaii to Tonga to Fiji in DX-12 game optimization ???

*scratches head*

What does DX12 have to do with whatever goes into the next PS?

 

[AtenRa]

*scratches head*

What does DX12 have to do with whatever goes into the next PS?

Im talking about Polaris

 

[zlatan]

Are you saying it will be worse than going from Hawaii to Tonga to Fiji in DX-12 game optimization ???

The PC games don't and won't ship shader binaries. With DXBC or SPIR-V the compliers will do their jobs. The VRAM management might be a problem, and sure there might be some programs where the new generation won't be faster, because the management is not optimised for the newest architectures. If this happens an application patch should fix the problem.

 

[AtenRa]

The PC games don't and won't ship shader binaries. With DXBC or SPIR-V the compliers will do their jobs. The VRAM management might be a problem, and sure there might be some programs where the new generation won't be faster, because the management is not optimised for the newest architectures. If this happens an application patch should fix the problem.

thanks

 

[airfathaaaaa]

The PC games don't and won't ship shader binaries. With DXBC or SPIR-V the compliers will do their jobs. The VRAM management might be a problem, and sure there might be some programs where the new generation won't be faster, because the management is not optimised for the newest architectures. If this happens an application patch should fix the problem.

i might start to think that its best to give the games to the gpu companies and let them optimise the game than to let any dev do it lol D:

 

[Sweepr]

Not sure if posted. Power consumption numbers for Polaris 10 and Polaris 11?

Nobody can say P10 better than Fiji.
This is P10 P11 technical development show scene photographed picture: In the "singularity ashes", all objects are equipped with light-emitting process alone, small to every hair laser cannon beams are independent sources, the extreme test DirectX 12 asynchronous computing power. Finally, there is the latest display technology features "AMD FreeSync" import "a low frame rate compensation (Low Framerate Compensation, LFC)," so that support FreeSync display, in the face of the panel frame rate lower than the minimum refresh rate conditions, avoid not smooth tabbing problem. What does this mean? P10 is the minimum FPS than FreeSync display update rate is low please hurry clever inferences P10 A black bar performance.

Single 6pin Additionally Pro Duo can be installed FirePro / Radeon dual drive GreenLand also later

Translation sux.

 

[Silverforce11]

Why do you guys set yourself up for such disappointment? It's pretty obvious from everything shown Vega will be Fury X's replacement, not Polaris 10.

Polaris 10 will target 390/X performance and at best it'll nip a the heels of the regular Fury or Nano at 1080/1440P.

The funny thing here is that Fury X is pretty crap for 1080p in general, so if Polaris manages to beat the 390X, that's REALLY close to Fury X.

It's bandwidth is going to hold it back at 4K for sure, but 1080 and 1440p, I think Polaris has a great chance to be 390X+ at ~120W.

 

[Silverforce11]

There are some situation where Polaris is incredibly fast. Faster than anything in the market. The secret is probably that special culling mechanism in the hardware, which helps the GPU to effectively cull those false positive primitives that aren't visible in the screen. Today's hardwares can't do this.
Single wavefront perfomance is also incredibly good. 10-100 times faster than anything in the market. This is good for VR.

I thought as much when they talked about the Primitive Discard Accelerator.

DICE has a primitive culling in Frostbite 3, using compute shaders to prepare the next frame, by discarding unseen objects. This is one of their major reason why the engine runs so damn fast compared to anything else on the market.

Polaris has this on hardware so it won't suffer shader latency and can do this for every game.

That's why they demo Hitman at the boat, because behind the visible boat are thousands of detailed NPCs, anyone with the game knows when you get close to the boat and look into it, the frame rate tanks. Polaris doesn't tank because it's not rendering stuff you can't see.

So in games with a lot of scene complexity, Polaris may well be Fury X+ performance. But overall, it won't. This is great for those who love higher MIN FPS though!

 

[Adored]

Not sure if posted. Power consumption numbers for Polaris 10 and Polaris 11?

Looks like Pitcairn and Cape Verde all over again. At 113W system draw it's much more like the 7850 than the 7870 so I'd expect that to be the cut-down P10.

If we go on the smart numbers I'd say 75W card draw for the P10 on the left and 40W card draw for the P11 on the right. The translation appears to suggest P10 has a 6-pin now but it'll be seen without one eventually.

It would be nice to know the performance but for both parts on show here I'd say P10 will be just under GTX 980 performance and P11 will be ~GTX 960.

They could both easily be Polaris 11 as well, just look at the 7750 and 7770 and the ~35W power draw between them, and we already know that P11 can run SWBF at around 80W system draw, at only 850MHz.

 

[Silverforce11]

Looks like Pitcairn and Cape Verde all over again. At 113W system draw it's much more like the 7850 than the 7870 so I'd expect that to be the cut-down P10.

Those are likely running with a frame rate limit though, AMD tends to do that in public demos.

In your next video, you should talk about my favorite new Polaris feature, the Primitive Discard Accelerator.

There's info there from DICE's GDC presentation where they talk about doing it via compute shaders. It's their secret sauce on why Frostbite games run so fast.

Edit:

For those interested:

http://www.frostbite.com/2016/03/optimizing-the-graphics-pipeline-with-compute/

There's several methods that DICE employs in their engine using compute. It takes some shader work to do, but the gains are worth it for them.

By not rendering at all unseen objects, which is different from normal Z-culling, which still renders it but skips the final output phase, you reduce the polygons processed per frame, you also reduce the bandwidth requirements per frame because objects not rendered at all, don't need texture maps, uv maps, shadow maps, etc.. and you reduce shader processing per frame for the same reasons.

It is a cheat engine in effect, and it's beautiful in theory, you have to wonder why GPUs didn't have this feature already. Imagine a game like The Division with all of its buildings and interiors, suddenly you are halving the scene complexity that your GPU has to render, because things unseen are discarded. The performance implications are huge for games with high scene complexity.

 

[Adored]

If run on a limiter you would expect P10 to have less power draw though (wider, lower clocks), or both to be closer to each other than this.

Assuming the same settings obviously, which it might not be. And obviously it might not even be P10 or P11. ;D

I'm looking to find more info on the primitive discard accelerator but it's not even available to devs yet, or if it is it just became available in the past couple of days.