Bring out Polaris now!

[thehotsung8701A]

I am stuck with a potato and as much as I love console gaming, I can't go ultrawide 1440p on console period.

 

[LTC8K6]

Aren't the first ones going to be on the low end of the performance scale?

 

[bononos]

Aren't the first ones going to be on the low end of the performance scale?

Is that right? I thought card manufacturers have been releasing only high end cards first.

 

[LTC8K6]

Is that right? I thought card manufacturers have been releasing only high end cards first.

Maxwell...?

750 and 750ti out first.

Roaring success...

 

[Stuka87]

nVidia always releases the top consumer card first, and then the big card 6-12 months later. Been this way for some time.

For AMD though, its different. Last node change top end cards all came many months before the mid range cards. It will most likely be the same thing year. With the releases spread out over a few months to get them all out.

 

[LTC8K6]

Yes, AMD should probably try something different this time.

I think the low end cards whet the appetite for the big cards.

 

[Elixer]

Yes, AMD should probably try something different this time.

I think the low end cards whet the appetite for the big cards.

Translation, our yields suck for the high end, let's use them for the low end instead!

All kidding aside, I do think that is both AMD's & nvidia's plan if yields suck for the high end.

I don't have a high confidence in GloFlo, even if they are using sammy tech.
Then again, TSMC don't have a stellar record on node shrinks either.

 

[Shmee]

Please note this is in the sub forum of AMD, nvidia talk should not be in here.

 

[ultima_trev]

I certainly wouldn't count on a Fiji successor this summer, most likely a 2017 release for the enthusiast-oriented chip.

Polaris 10 is a small chip with GTX 950 like performance but in a smaller power envelope, per what AMD demoed at CES anyhow.

Polaris 11 is supposedly a 232 mm^2 mainstream chip (Pitcairn successor), which will probably perform between 380X and 390 but offer much better power efficiency than either.

 

[3DVagabond]

I certainly wouldn't count on a Fiji successor this summer, most likely a 2017 release for the enthusiast-oriented chip.

Polaris 10 is a small chip with GTX 950 like performance but in a smaller power envelope, per what AMD demoed at CES anyhow.

Polaris 11 is supposedly a 232 mm^2 mainstream chip (Pitcairn successor), which will probably perform between 380X and 390 but offer much better power efficiency than either.

Polaris 10 was running @ 850MHz and less than 50w. That's very low. I don't think we saw it's true potential.

 

[ultima_trev]

Indeed. ~40-50 watts Sounds par for the course for the laptop version of Polaris 10, while the desktop version will most likely be subject to higher clocks and power envelopes. I wouldn't be surprised if it (a mainstream/entry part) edged out a GTX 960 and once again extended the gap between IGP and low end discreet cards.

 

[Silverforce11]

Polaris 10 was running @ 850MHz and less than 50w. That's very low. I don't think we saw it's true potential.

On paper, a ~120mm2 Polaris 10 with x2.4 density as 28nm, is equivalent to GCN 1.2 of ~288mm2 on 28nm.

This is a bit smaller than Tonga.

But one has to account for the new uarch gains. It could match the 380X. But pretty much it requires the node to perform as advertised as well as some decent uarch improvements.

 

[Samwell]

On paper, a ~120mm2 Polaris 10 with x2.4 density as 28nm, is equivalent to GCN 1.2 of ~288mm2 on 28nm.

This is a bit smaller than Tonga.

But one has to account for the new uarch gains. It could match the 380X. But pretty much it requires the node to perform as advertised as well as some decent uarch improvements.

Where are these 2,4x coming from?
uarch gains don't help you much in perf/transistor. They help you more in perf/w. You add new features, more power gating and so on, all stuff which costs transistors. Just look at Kepler vs Maxwell. GK104 vs GM204 is 50% more Transistors for 60% more perf. But 2 vs 4 gb ram. So perf/transistor probably stayed the same and i wouldn't expect different for polaris.

Polaris 10 was clocked at 850mhz matching a GTX950, i think 1100mhz sounds realistic which is 30% more speed. About on par with GTX960 with lower than 75W TDP would be my bet for desktop.

 

[raghu78]

Where are these 2,4x coming from?
uarch gains don't help you much in perf/transistor. They help you more in perf/w. You add new features, more power gating and so on, all stuff which costs transistors. Just look at Kepler vs Maxwell. GK104 vs GM204 is 50% more Transistors for 60% more perf. But 2 vs 4 gb ram. So perf/transistor probably stayed the same and i wouldn't expect different for polaris.

Polaris 10 was clocked at 850mhz matching a GTX950, i think 1100mhz sounds realistic which is 30% more speed. About on par with GTX960 with lower than 75W TDP would be my bet for desktop.

Performance never scales perfectly linearly by just adding transistors. btw the GM204 (5.2 billion transistors at 398 sq mm) beats GK110 (7.1 billion transistors at 550 sq mm). Even if we account for the fact that GK110 has FP64 that still does not explain how a smaller GPU with almost 2 billion transistors less can beat the larger GPU. Its well understood that microarchitectural improvements help increase perf/watt, perf/transistor and perf/sq mm.

https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_980/26.html

https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_980/27.html

This is the AMD forum. If you don,t stop discussing Nvidia here, I will infract all of you doing it.
This is your only warning
Markfw900

 

[Samwell]

Just look at the frequency, 30% difference, than you have your answer and like you said you don't know how many transistors were needed for the gpgpu stuff. That's why i took 2 more comparable gpus. But this is getting too offtopic for the amd section. Just wanted to show that in my opinion perf/ transistor stays about the same if frequency doesn't go up substancially (Which could also happen, no one knows how high finfet can go)

 

[cbn]

Is that right? I thought card manufacturers have been releasing only high end cards first.

Maxwell...?

750 and 750ti out first.

Roaring success...

Don't forget the HD 4770 as well.

It was AMD's first 40nm card.

 

[Sonikku]

Yes, AMD should probably try something different this time.

I think the low end cards whet the appetite for the big cards.

Bring on the low end! The 7770 is getting too long in the tooth...

 

[12age]

gahhhhh I wonder if I should sell my 980 while its still got value and build a full AMD system. I wonder if having a zen proc and Polaris GPU will have any difference.

 

[Doom2pro]

Turns out Polaris 10 is the mid range, Polaris 11 the smallest and Vega the big boy.

 

[flopper]

couple of months to go or so thats soon...