1080 = new 680, Polaris 10 = new 7870?

[NTMBK]

Is it just me, or are both AMD and NVidia reusing the 28nm playbook? The 680 was slightly faster than the 580 it replaced, with much better efficiency. Same story as the 1080. 7870 was slower, but a great perf/$ card which served people well for years. Polaris 10 will fit right into that niche, according to rumours.

Only missing piece is a 7970 equivalent.

 

[maddie]

Is it just me, or are both AMD and NVidia reusing the 28nm playbook? The 680 was slightly faster than the 580 it replaced, with much better efficiency. Same story as the 1080. 7870 was slower, but a great perf/$ card which served people well for years. Polaris 10 will fit right into that niche, according to rumours.

Only missing piece is a 7970 equivalent.

Yes.

A rather big piece. Came out first. Topped the charts. Top $. Yes, a rather big piece.

 

[Erenhardt]

P10 should have the die size around the same as 7870 <- amazing chip besting maxwells gtx960 in recent titles in both performance and performance/watt, and it is even smaller than 960.

7870 is a 120W card. Around what we expect from P10, maybe even on the downside. Even when we assume P10 will take the same amount of juice as 7870, and we apply the 2.5x efficiency on top of it we arrive at:

2.5x106%=265% right where the 980ti/tx is.

If we assume P10 is really a 150W card, as rumored, 30 additional watts is 25% more juice, at 2.5x efficiency, 1.25x2.5x106%=331% of relative performance - somewhere off the charts more than 25% faster than 980ti.

As for 7970 there will be Vega. It will probably feature 2 chips, like polaris. One would be around 350-400mmsq as a 7970 counterpart, the other will be hawaii/fiji replacement at around 500mmsq.

I think amd will focus now on small chips and perf/mmsq so they can later use with interposer in die-stacking-single-GPU Navi era as well as profitable gaming APUs.

 

[vissarix]

Is it just me, or are both AMD and NVidia reusing the 28nm playbook? The 680 was slightly faster than the 580 it replaced, with much better efficiency. Same story as the 1080. 7870 was slower, but a great perf/$ card which served people well for years. Polaris 10 will fit right into that niche, according to rumours.

Only missing piece is a 7970 equivalent.

lol you have mixed things up a bit ... the 7970 when launched was 30% faster then the gtx580, and then after few months the gtx680 comes out beating the 7970 while being more power efficient...amd released a ghz version of the 7970 just to get the performance crown back...

so i wouldnt say the gtx680 was slightly faster vs the gtx580...+30% faster...

 

[Erenhardt]

lol you have mixed things up a bit ... the 7970 when launched was 30% faster then the gtx580, and then after few months the gtx680 comes out beating the 7970 while being more power efficient...amd released a ghz version of the 7970 just to get the performance crown back...

so i wouldnt say the gtx680 was slightly faster vs the gtx580...+30% faster...

at short burst 720p benchmarks, maybe.

 

[Kuiva maa]

at short burst 720p benchmarks, maybe.

Unless I am mistaken , Polaris 10 is 232mm^2 of Samsung 14nmFF which roughly amounts to ~250mm^2 of tsmc 16nmFF. Pitcairn is 212 and GK104 was 294,with GP104 being ~295mm^2 , no ? So Polaris 10 will be noticeably bigger than Pitcairn in analogy . Where it will sit performance-wise depends on arch features and clock speed. I think in DX12 it might be very competitive vs GTX1080. And in general Polaris 10 should clearly beat fury X/Titan X/980ti. If a cut down 1080 can reportedly do that in the form of 1070, then Polaris will probably do it too.

 

[digitaldurandal]

Unless I am mistaken , Polaris 10 is 232mm^2 of Samsung 14nmFF which roughly amounts to ~250mm^2 of tsmc 16nmFF. Pitcairn is 212 and GK104 was 294,with GP104 being ~295mm^2 , no ? So Polaris 10 will be noticeably bigger than Pitcairn in analogy . Where it will sit performance-wise depends on arch features and clock speed. I think in DX12 it might be very competitive vs GTX1080. And in general Polaris 10 should clearly beat fury X/Titan X/980ti. If a cut down 1080 can reportedly do that in the form of 1070, then Polaris will probably do it too.

I think it will trade blows with 390x at best. Waiting for Vega.

 

[airfathaaaaa]

I think it will trade blows with 390x at best. Waiting for Vega.

its bigger than a 390x......almost 30mm2 bigger..

 

[el etro]

Polaris 10 measuring 232mm2 of 14LPP equals something north of 250mm2 of TSMC, if Polaris really have the rumored perf/mm2 lead over Pascal so is better to expect P10 to be faster than it looks.
Also the best thing would be if AMD achieves the perf/w lead, it would take them to build cheaper cards and have the cost advantage over Nvidia. See GTX960/GM206 strategy as a example, a GPU/card that is ultimately pretty cheap to manufacture, and was intended to substitute the expensive GK104 cards.

 

[Redentor]

Then NVIDIA fanboys will say: we don't care about power consumption, we want performances! LOL


Threadcrapping and trolling are not allowed
Markfw900

 

[Kuiva maa]

I think it will trade blows with 390x at best. Waiting for Vega.

I find this highly unlikely. Koduri tweeted this:



"@Gchip Polaris is 2.5 times brighter today than when Ptolemy observed it in 169 A.D"

...which probably is a reference to density? Then there was the other perhaps a bit far fetched hint from Hook :

"@GFXChipTweeter As I've told you 16 million times, nothing less than the 8-methyl-N-vanillyl-6-nonenamide-infused Carolina Reaper will do..."

People made the hypothesis that he hinted of polaris 10 having 8.6 billion transistors which means Fiji class (probably faster due to newer arch). No idea what 16million stands for though.

 

[Tuna-Fish]

"@GFXChipTweeter As I've told you 16 million times, nothing less than the 8-methyl-N-vanillyl-6-nonenamide-infused Carolina Reaper will do..."

People made the hypothesis that he hinted of polaris 10 having 8.6 billion transistors

And that was just straight up stupid. 8-methyl-N-vanillyl-6-noenamide is capsaicin.

 

[airfathaaaaa]

I find this highly unlikely. Koduri tweeted this:



"@Gchip Polaris is 2.5 times brighter today than when Ptolemy observed it in 169 A.D"

...which probably is a reference to density? Then there was the other perhaps a bit far fetched hint from Hook :

"@GFXChipTweeter As I've told you 16 million times, nothing less than the 8-methyl-N-vanillyl-6-nonenamide-infused Carolina Reaper will do..."

People made the hypothesis that he hinted of polaris 10 having 8.6 billion transistors which means Fiji class (probably faster due to newer arch). No idea what 16million stands for though.

https://en.wikipedia.org/wiki/Capsaicin
i just hope that he wasnt reffering to the heat lol

 

[R0H1T]

The Polaris 10 has to have at least 8 billion transistors, probably closer to nine, if it's going to beat 1070 decisively.

 

[Slaughterem]

I find this highly unlikely. Koduri tweeted this:



"@Gchip Polaris is 2.5 times brighter today than when Ptolemy observed it in 169 A.D"

...which probably is a reference to density? Then there was the other perhaps a bit far fetched hint from Hook :

"@GFXChipTweeter As I've told you 16 million times, nothing less than the 8-methyl-N-vanillyl-6-nonenamide-infused Carolina Reaper will do..."

People made the hypothesis that he hinted of polaris 10 having 8.6 billion transistors which means Fiji class (probably faster due to newer arch). No idea what 16million stands for though.

16 million refers to the scoville heat unit of the carolina reaper which averages 15.7 and has reached 2.2 million on the scale. This could have been a clue for the frequency of Polaris arch.
https://en.wikipedia.org/wiki/Carolina_Reaper

 

[tential]

Nvidia should have opened founders edition preorders on announcement. Polaris would have been crushed right away. I'm positive next gen thats how they one up what they did now

 

[maddie]

The Polaris 10 has to have at least 8 billion transistors, probably closer to nine, if it's going to beat 1070 decisively.

Beat 1070 decisively. Are you serious?

Is this what some are saying? I consider my expectations to be at the upper end of the predicted performance range, using a reasoned argument, but this is ridiculous.

If it beats 1070 decisively, you're saying in effect, that it is around a 1080.

Polaris 10 should be able to match 1070.

 

[Slaughterem]

Nvidia should have opened founders edition preorders on announcement. Polaris would have been crushed right away. I'm positive next gen thats how they one up what they did now

Please explain how the Pascal 1080 and 1070 crushes the Polaris offering from the recent marketing announcement.

 

[R0H1T]

16 million refers to the scoville heat unit of the carolina reaper which averages 15.7 and has reached 2.2 million on the scale. This could have been a clue for the frequency of Polaris arch.
https://en.wikipedia.org/wiki/Carolina_Reaper

The reaper is averaging around 2 million, pure Capsaicin is 16 million. Resiniferatoxin is a whopping 16 Billion units on the Scoville scale D:

 

[R0H1T]

Beat 1070 decisively. Are you serious?

Is this what some are saying? I consider my expectations to be at the upper end of the predicted performance range, using a reasoned argument, but this is ridiculous.

If it beats 1070 decisively, you're saying in effect, that it is around a 1080.

Polaris 10 should be able to match 1070.

A miners edition with GDDR5x could do it (vs stock 1070) & yes I am serious but it depends largely on the clocks we'll see on full Polaris 10.

 

[maddie]

Nvidia should have opened founders edition preorders on announcement. Polaris would have been crushed right away. I'm positive next gen thats how they one up what they did now

You realize that to crush AMD and Polaris you must have stock in sufficient quantities to satisfy all demand? This is assuming that Polaris is non-competitive. Unless you have some insider knowledge and not sharing.

Do you really expect us to believe that Nvidia is rolling in stock ready to saturate the retailers? What a joke.

Unless you expect Nvidia to sell all of the 1070 cards they will ever sell and satisfy the market for this new card in June, a market will still exist. I wonder what they will produce after that miraculous feat? I guess that will release so many additional wafers for GP100 that we'll get big Pascal earlier.

Normally your posts are well reasoned, but the last couple days since the unveiling of 1080 & 1070 has seen some very, very strange posts by you. Very poorly thought out and reasoned positions. Almost as if a different person has emerged.

 

[Slaughterem]

The reaper is averaging around 2 million, pure Capsaicin is 16 million. Resiniferatoxin is a whopping 16 Billion units on the Scoville scale D:

You are correct 1.57 up to 2.2 million - point being it could allude to freq range of the arch in GHz

 

[R0H1T]

You are correct 1.57 up to 2.2 million - point being it could allude to freq range of the arch in GHz

I think those hints were just for the Pro Duo, though I'd be happy to be proven wrong.

 

[renderstate]

The way some are normalizing Polaris 14nm area to 390 28nm area is not correct and leads to overoptimistic numbers. Not every unit of the GPU is going to scale linearly with process. In particular the physical I/O interfaces don't scale well and sometime they can actually scale negatively to compensate for different process characteristics (and inexperience..). At this point is more likely that Polaris matches a 390x in terms of transistor count, or perhaps it's s bit less, but it makes up for it with higher efficiency and better process.


Sent from my iPhone using Tapatalk

 

[Despoiler]

I find this highly unlikely. Koduri tweeted this:



"@Gchip Polaris is 2.5 times brighter today than when Ptolemy observed it in 169 A.D"

...which probably is a reference to density? Then there was the other perhaps a bit far fetched hint from Hook :

"@GFXChipTweeter As I've told you 16 million times, nothing less than the 8-methyl-N-vanillyl-6-nonenamide-infused Carolina Reaper will do..."

People made the hypothesis that he hinted of polaris 10 having 8.6 billion transistors which means Fiji class (probably faster due to newer arch). No idea what 16million stands for though.

8.6 and 16 was in reference to Radeon Pro Duo, which is already released.