AMD Polaris architecture to succeed Graphics Core Next

Polaris article from Wccftech with "new leaks", treat all as rumor:

Polaris 11 [Baffin]
target - 1080p60 gaming for desktop & laptops
2 GPUs
16CU 1024 SPs & 20CU 1280 SPs
4GB of GDDR5/X
~50W


Polaris 10 [Ellesmere]
target - 1440p60 DX12 gaming for desktop
2 GPUs
36CU 2304 SPs & 40CU 2560 SPs
8GB GDDR5/X
~100-110W


Vega 10 [release in early 2017]
target - successor of Fury line
64CUs 4096 SPs
16GB HBM2
unknown wattage


http://wccftech.com/amd-radeon-r9-480-470-polaris-10-polaris-11/
 
Polaris article from Wccftech with "new leaks", treat all as rumor:

Polaris 11 [Baffin]
target - 1080p60 gaming for desktop & laptops
2 GPUs
16CU 1024 SPs & 20CU 1280 SPs
4GB of GDDR5/X
~50W


Polaris 10 [Ellesmere]
target - 1440p60 DX12 gaming for desktop
2 GPUs
36CU 2304 SPs & 40CU 2560 SPs
8GB GDDR5/X
~100-110W


Vega 10 [release in early 2017]
target - successor of Fury line
64CUs 4096 SPs
16GB HBM2
unknown wattage


http://wccftech.com/amd-radeon-r9-480-470-polaris-10-polaris-11/

If true, I really wonder how far they could push the power of a Gpu like the 750 ti that basically anyone could easily use to turn their pc into a gaming one. Polaris 11 would make a perfect Gpu for that, if the info is true.
 
This is the first time I'm really considering switching to AMD especially with memory reports for the Nvidia Pascal cards and the spotty DX12 performance compared to AMD.
 
A 14Gbps GDDR5X on a 512 bit bus will be able to achieve 896GB/s of bandwidth.
A top speced HBM2 on a 4096 bit bus (which seems to be a typical config for HBM chips as both Fiji and GP100 are using it) is hitting 1000GB/s.
The difference isn't that big as you make it sound really. Give or take several MC wrinkles here and there, some savings in GDDR5X, some issues in HBM2 and you'll end up with pretty much the same bandwidth.

I can't imagine either AMD or Nivida using 512 bit busses on GDDR5X (and given chip availability I think that would be only feasible with 16GB of it). Currently manufacturers are advertising their early GDDR5X as running up to 12GT/s, which would put bandwidth at 384GB/s for a 8GB 256-bit bus. Which is plenty for all but the highest-end cards, anyway. It's not like the 980Ti is being horrendously crippled with 336GB/s.
 
... and the spotty DX12 performance compared to AMD.

so good


75c2d5ad_ki_1920.jpeg


b220b12f_ki_2560.jpeg


9fd858da_ki_3840.jpeg



http://gamegpu.com/action-/-fps-/-tps/killer-instinct-test-gpu


61718521.jpg
 
To me it is too obvious and perfect that the leaks line up nicely as semi-custom parts: Polaris 11 = NX and Polaris 10 = PS4K.

Polaris 10 [Ellesmere]
target - 1440p60 DX12 gaming for desktop
36CU 2304 SPs & 40CU 2560 SPs
~100-110W

Exactly 2X PS4 GPU for the 36CU version

Polaris 11 [Baffin]
target - 1080p60 gaming for desktop & laptops
2 GPUs
16CU 1024 SPs & 20CU 1280 SPs
4GB of GDDR5/X
~50W

16CU+new architecture = bit more powerful than PS4 plus nice and low powered.
 
To me it is too obvious and perfect that the leaks line up nicely as semi-custom parts: Polaris 11 = NX and Polaris 10 = PS4K.

Polaris 10 [Ellesmere]
target - 1440p60 DX12 gaming for desktop
36CU 2304 SPs & 40CU 2560 SPs
~100-110W

Exactly 2X PS4 GPU for the 36CU version

Polaris 11 [Baffin]
target - 1080p60 gaming for desktop & laptops
2 GPUs
16CU 1024 SPs & 20CU 1280 SPs
4GB of GDDR5/X
~50W

16CU+new architecture = bit more powerful than PS4 plus nice and low powered.

Err, is this a joke?
 
To me it is too obvious and perfect that the leaks line up nicely as semi-custom parts: Polaris 11 = NX and Polaris 10 = PS4K.

Polaris 10 [Ellesmere]
target - 1440p60 DX12 gaming for desktop
36CU 2304 SPs & 40CU 2560 SPs
~100-110W

Exactly 2X PS4 GPU for the 36CU version

Polaris 11 [Baffin]
target - 1080p60 gaming for desktop & laptops
2 GPUs
16CU 1024 SPs & 20CU 1280 SPs
4GB of GDDR5/X
~50W

16CU+new architecture = bit more powerful than PS4 plus nice and low powered.

They're not gonna put a 350$-550$ GPU in a console.
 
That moment knowing you put up a bad example of testing dx12 improvements and not one 390 or 390x is used in any of the tests.

That moment when you realize so far anything running strictly on Windows 10 store app, does not equal to a great testing example. Especially when it's dx11 game.

come one. You can't play with "ideology" of buts and ifs. You play with frames per second, and those right there are the performance you get and they are damn sweet :)

290x on 980ti tail
 
They're not gonna put a 350$-550$ GPU in a console.

Don't confuse retail price of PC GPU cards with what is used in consoles [just a GPU module]. In PC land we pay for GPU chip, lots of VRAM, PCB, power delivery chips, ports, cooling, and of course, we pay for a nice profit chunk.
 
Why going so low ? Should have put a 200w card to blow away the Fury.

Thats the only piece of the puzzle left. They said Polaris will address as many market segments as possible, but if Polaris 10 is the R9 480(X) then they're leaving the high end/enthusiast out. Maybe they just want to cover the $150-350 price tiers and are willing to let Nvidia have the $500+ market. Which would make sense why Nvidia is releasing three GP104 cards instead of two (if the rumors are true).
 
Don't confuse retail price of PC GPU cards with what is used in consoles [just a GPU module]. In PC land we pay for GPU chip, lots of VRAM, PCB, power delivery chips, ports, cooling, and of course, we pay for a nice profit chunk.

Got it. Still wouldn't that mean that the PS4k would be based on a super modern high end PC GPU chip? I mean that would be nice but it seems unlikely to me.
 
So it isn't.

Ellesmere is way more than 2x the power of a PS4.

Sorry I see what you mean. Reading OsirisBlack's leak thread and he talked in non-technical terms about the PS4K being 2X GPU and much faster etc. I'm thinking non-technical people were sharing paper specs like 2304SP (PS4=1152SP) and faster clocks of 1000Mhz+ (PS4=800Mhz)

Zojirushi said:
They're not gonna put a 350$-550$ GPU in a console.

I don't think they would start that high. The article suggests AMD may be aiming <$300. Maybe $250 for the lower end part? HD7850/70 were $250/$350 at launch albeit 18/20 months before PS4 launch.
 
Have you looked at the predicted die size of that GPU? There is no way Sony is going to be paying for an APU that fits a gpu that large.
Polaris 10 is ~230mm2, no?

Ellesmere is way more than 2x the power of a PS4. A PS4K is looking at something like a 280x.
I guess PS4K will have 2 CUs for redundancy like PS4.

Looks like 2x the PS4 power.

1.84TF vs ~3.5TF
 
I may as well stretch this theory to breaking point.....

Ellesmere Port is a town just outside Liverpool, the codename of the PS4 APU.
 
Sorry I see what you mean. Reading OsirisBlack's leak thread and he talked in non-technical terms about the PS4K being 2X GPU and much faster etc. I'm thinking non-technical people were sharing paper specs like 2304SP (PS4=1152SP) and faster clocks of 1000Mhz+ (PS4=800Mhz)

Ah. No, it is far more likely to be 2x the absolute capabilities of the ps4 gpu, not necessarily the SPs or Clock speed.
 
From whatever was leaked Polaris looks kinda interesting. Either the performance expectations created by AMD demonstrations are overblown or Polaris will finally provide the graphics efficiency improvements which GCN should've got years ago. The third possibility would be a change of number of SIMDs in a CU which may also improve relative gaming performance but will probably net a negative effect on compute capability of the architecture.

Since we're like up to our heads in the silly season it's rather pointless to even follow these rumors now as they are all over the place at the moment. I'm just waiting for the info I can definitely trust to be correct.
 
PS4 APU is 348mm2 of which Jaguar is 50-60mm2? and on 28nm.

Sure, The new CPU is liable to have more transistors too, unless they opt to go with very high clock speeds on jaguar, which they can't as that would destroy their power budget or TDP.

One also has to remember that large dies are likely to be cheaper on 28nm due to the maturity of the process. A die of similar size at 16nm is liable to be much more expensive.
 
So Polaris 10 is basically a 100w Fury X, sounds like they managed that 2.5x performance per watt improvement after all.

huh? doesn't the fury x have 4000SPs?


Polaris 10 is ~230mm2, no?


I guess PS4K will have 2 CUs for redundancy like PS4.

Looks like 2x the PS4 power.

1.84TF vs ~3.5TF

and 230mm^2 is tiny, the gtx 970 by comparison is a 398mm^2 die, the 980 ti a 600mm^2 die

edit: hd7850(same gpu in ps4 basically) had a 212mm2 die

polaris should be dirt dirt dirt cheap, though I'm worried about the performance if polaris 10 really has such a tiny die.
 
polaris should be dirt dirt dirt cheap, though I'm worried about the performance if polaris 10 really has such a tiny die.

Depends on yields of the new process node. They might have gone with small dies purely due to that.

Apple has been using them already, for fairly sizable dies. However, Apple is probably more willing to swallow lower yields on account of their profit margins.
 
It still has to fit interfacing logic and 8 X86 cores on the same die.
That makes it smaller than actual PS4's APU or close to the same size.

8x Jaguar cores is less than 80mm2 in 28nm.

Polaris 10 APU for PS4k can end being 300-320mm2... that won't be that expensive compared with PS4's APU unless the yelds are really bad.
 
Performance wise, Polaris 10 seems to be comparable with the Fury X according to the Capsaicin demonstration and these latest rumors.

If they can match a 4000 sp fury x with a 2500 sp card with only 100w power consumption that would be quite extraordinary

That would mean that if they ever made big die card it would be an unholy monster performance wise

I would love for this to be true, but it seems way too good to be true, it would mean their new architecture is an enormous improvement over GCN and doesn't every rumor say polaris is just a new gcn iteration.
 
Performance wise, Polaris 10 seems to be comparable with the Fury X according to the Capsaicin demonstration and these latest rumors.

We can't know the true performance of Polaris chips just yet. Recent demos include:

- smaller Polaris 11 rendered "4K VR content" while being passively cooled. :)
https://twitter.com/ryanshrout/status/709761872778514433

- bigger Polaris 10 got demoed on Hitman on unknown visual settings, claiming it reached 1440p60.
 
If they can match a 4000 sp fury x with a 2500 sp card with only 100w power consumption that would be quite extraordinary

That would mean that if they ever made big die card it would be an unholy monster performance wise

I would love for this to be true, but it seems way too good to be true, it would mean their new architecture is an enormous improvement over GCN and doesn't every rumor say polaris is just a new gcn iteration.
I guess clock was a lot to do with that.

But I agree with you it is really unlicky to happen... I expect 10-20% increase in performance per SP at same clock.
 
I guess clock was a lot to do with that.

But I agree with you it is really unlicky to happen... I expect 10-20% increase in performance per SP at same clock.

Hm I guess now that you mention clockspeed it doesn't sound that unreasonable afterall... maybe

20 percent faster SP at the same clock and 30 percent higher clockspeed (I had no idea the furyx only has a 1ghz core clock, most nvidia gpus go up to 1400mhz easily with an OC) would get you into that ballpark.

2560*1.2*1.3 = 3990 fury x sp's worth of performance

We'll see I guess, it sure would be super sweet if they could pull that off, especially at <130w TDP

I'm drooling at the thought of what a 600mm2 vega card in 2017 would do if they can pull off that kind of effiency and performance
 
Baffin looks like something Nintendo would like. Low Wattage, nice bang for your buck. Remember the entire power usage of the WiiU is about 36 watts so a 50 watts GPU + a possible ARM chip would be right up their street.
 
Baffin looks like something Nintendo would like. Low Wattage, nice bang for your buck. Remember the entire power usage of the WiiU is about 36 watts so a 50 watts GPU + a possible ARM chip would be right up their street.

This chip plus larger mobo, more GDDR5, HDD, DVD drive, LAN & WiFi modules, etc. Nice console for <100W.
 
Liking the way these rumors are coming together. Ready to see this stuff announced with prices so I can make a buying decision.
 
Polaris 10 is ~230mm2, no?


I guess PS4K will have 2 CUs for redundancy like PS4.

Looks like 2x the PS4 power.

1.84TF vs ~3.5TF

PS4K would be Polaris 11 not 10 if it is 2x PS4.

Until Phil Spencer recent comments I thought MS would launch 2 Skus this year.

Xbox slim 8 GB QDR and Polaris 11 for 1080p60 gaming @ $299.

And

Xbox pro 12 - 16 GB of QDR Polaris 10 for 1440 / 60 or 4K/30 @ $499
 
http://www.planet3dnow.de/cms/23084-polaris-mit-bis-zu-1154-mhz-und-neuer-stromsparfunktion/

Some more info from a Linux patch. It basically confirms the memory configuration, clock ranges and some new power saving feature. I wonder if it's related to a recent AMD patent that was generating a bit of a buzz.

From Google Translate:

In Ellesmere 8 channels and 8 banks of 32 bits can be specified, Baffin is designed to provide only 4 channels, but can be fitted with 8 banks. From the specifications can be seen that AMD will initially continue to rely on GDDR5 memory. With four channels logically produces a 256-bit memory interface, eight channels of 512 bits. In both cases, a maximum memory configuration of 8 GiB GDDR5. As maximum clock a very high value of 6000 MHz is specified in each case. This corresponds to the effective clock speed of the GDDR5 memory and is already in models like AMD's Radeon R9 380 or GeForce GTX see 960th To what extent the currently specified maximum clock speed of 1154 MHz, which is valid for both chips, is also so incorporated into the series remains to be seen.
 
come one. You can't play with "ideology" of buts and ifs. You play with frames per second, and those right there are the performance you get and they are damn sweet :)

290x on 980ti tail

AMD right now in dx12 beats a 980gtx and in most cases on higher resolutions a 980Ti.

But that game is a poor test plain and simple because your using DX11.

I also love the fact you didn't include any test's with any AMD 8gb card which there are quite a few in the 390, 390x config. You also didn't include the R9 fury, or nano.
 
Top Bottom