AMD Polaris architecture – GCN 4.0

[Mauller]

Mildly disappointing if this turns out to be the biggest chip we get for the first round. Was hoping to see something in the 290-300mm2 class, since that would reasonably guarantee being 980 TI+ performance.

You have to consider that it is a die shrink and using finfett. so a 200mm^2 14nm design would more than likely be over 300mm^2 on 22nm

 

[Orangey]

Anyone remember the linkedin die shrink design for the consoles that never materialized?

 

[bru]

It looks like AMD have a 232MM2 GPU in the works:

http://videocardz.com/58237/amds-project-f-is-232mm2-discrete-gpu-made-in-14lpp-process

If that is the larger Polaris 11 I suspect that it probably is around GTX970/R9 290 level performance,but probably runs off a single PCI-E power connector.



Not when you paid under £110 for it!!!

Well I really do hope that isn't right or it isn't a GPU.

The card that Raja first held up to the press at CES was supposedly smaller than Cape-Verde 123mm^, so if this is the bigger Polaris and it is only 232mm^ then, that isn't going to get the job done.

There is no way that a GPU that small can be that efficient and be that powerful and if AMD think that a new top drawer chip for this year being slower than their currant top chip and having more memory is a good idea, then wow, really.

The only way that a chip that small could be faster then the FuryX is if, as Some of us said many, many months ago the performance from this new 14/16nm node is two maybe two and a half times the currant performance. When we said this we were shouted down saying how stupid and ridiculous that idea was.

I really do pray to god that this isn't the bigger of the two Polaris's that are coming sometime before September.

 

[Mauller]

Well I really do hope that isn't right or it isn't a GPU.

The card that Raja first held up to the press at CES was supposedly smaller than Cape-Verde 123mm^, so if this is the bigger Polaris and it is only 232mm^ then, that isn't going to get the job done.

Here are some quotes from Samsungs website.

14LPE (Early edition) targets the early technology leaders and time-to-market customers such as mobile application SoCs to meet the latest mobile gadgets’ aggressive schedule and improved performance/power requirements. 14LPE is the first foundry process technology manufactured in the foundry industry with the successful volume ramp-up. 14LPE offers 40% faster performance; 60% less power consumption; and, 50% smaller chip area scaling as compared to its 28LPP process.

14LPP (Performance boosted edition) is the 2nd FinFET generation which the performance is enhanced up to 10%. 14LPP is the single platform for every application designs with the improved performance for computing/Network designs and the lowered power consumption for Mobile/Consumer designs. 14LPP will be the main process technology offering in 2016 and after.

the key phrase being 50% smaller chip area scaling as compared to its 28LPP process

So you can consider a 14nm chip of 232mm^2 to be comparable to a 464mm^2 22nm planar chip. Although that is just the size savings and discounting other metrics that will improve performance. That is just process performance, discounting the increased performance from architectural improvements that save power and improve performance.

So it is still feasible that this 232mm^2 chip is the 290x - 980ti performance chip.

Edit: As another note, the 290X is also only 438MM^2 in size, with the 980ti being 600mm^2 in size. So you could consider the 232mm^2 chip almost a direct shrink of the 290X

We also know nothing about performance metrics, this chip could more than likely clock up to 1.5ghz with ease, maybe nearing 2ghz boost. Keeping around the 300Watt power limit but greatly improving performance for that power limit.

 

[AllBodies]

Well I really do hope that isn't right or it isn't a GPU.

The card that Raja first held up to the press at CES was supposedly smaller than Cape-Verde 123mm^, so if this is the bigger Polaris and it is only 232mm^ then, that isn't going to get the job done.

There is no way that a GPU that small can be that efficient and be that powerful and if AMD think that a new top drawer chip for this year being slower than their currant top chip and having more memory is a good idea, then wow, really.

The only way that a chip that small could be faster then the FuryX is if, as Some of us said many, many months ago the performance from this new 14/16nm node is two maybe two and a half times the currant performance. When we said this we were shouted down saying how stupid and ridiculous that idea was.

I really do pray to god that this isn't the bigger of the two Polaris's that are coming sometime before September.

Fingers crossed this is actually a very high-end laptop chip which will come later and sit in the mid-range for desktop. And Polaris 11 is more like 290-300mm2.

The GTX 980m is 398mm2, so 232mm2 on 14nm would kick the crap out of that.

 

[humbug]

Well I really do hope that isn't right or it isn't a GPU.

The card that Raja first held up to the press at CES was supposedly smaller than Cape-Verde 123mm^, so if this is the bigger Polaris and it is only 232mm^ then, that isn't going to get the job done.

There is no way that a GPU that small can be that efficient and be that powerful and if AMD think that a new top drawer chip for this year being slower than their currant top chip and having more memory is a good idea, then wow, really.

The only way that a chip that small could be faster then the FuryX is if, as Some of us said many, many months ago the performance from this new 14/16nm node is two maybe two and a half times the currant performance. When we said this we were shouted down saying how stupid and ridiculous that idea was.

I really do pray to god that this isn't the bigger of the two Polaris's that are coming sometime before September.

There is another way to look at its, 14nm is 2x the transistor density.

The Fury-X is 550mm and has 8.9bn Transistors.

2x 232mm = 464mm = 82% of Fury-X / 7.5bn Transistors.

The 290X is 432mm with 6.2bn Transistors.

So that speculation would have it between a 290X and a Fury-X, or 980 and 980TI. (this without any architectural performance improvement)

Apparently the architecture has been completely reworked and has a performance improvement over GCN 1.2.

So, its possible. i suspect its what they are shooting for with it.

Edit: Mauller beat me to it.

 

[AllBodies]

There is another way to look at its, 14nm is 2x the transistor density.

The Fury-X is 550mm and has 8.9bn Transistors.

2x 232mm = 464mm = 82% of Fury-X / 7.5bn Transistors.

The 290X is 432mm with 6.2bn Transistors.

So that speculation would have it between a 290X and a Fury-X, or 980 and 980TI. (this without any architectural performance improvement)

Apparently the architecture has been completely reworked and has a performance improvement over GCN 1.2.

So, its possible. i suspect its what they are shooting for with it.

To add to this I found a source claiming GloFo/Samsung 14LPP is actually 15% smaller than TSMC 16FF+. And I'm sure I've heard this elsewhere before too, that Samsung's process is mildly smaller.

So given that TSMC's is 2x the density over 28nm, the AMD cards will have ~2.3x.

The 232mm2 Polaris could have as much as 8 Billion if this is the case.

 

[humbug]

To add to this I found a source claiming GloFo/Samsung 14LPP is actually 15% smaller than TSMC 16FF+. And I'm sure I've heard this elsewhere before too, that Samsung's process is mildly smaller.

So given that TSMC's is 2x the density over 28nm, the AMD cards will have ~2.3x.

The 232mm2 Polaris could have as much as 8 Billion if this is the case.

Right so the Transistor deficit from the Fury-X is only 11%, the faster HBM2 buffer and architectural improvements should easily make that up.

I wouldn't be at all surprised if this thing is as fast as a Fury-X / GTX 980TI or perhaps a little faster. If they can pack almost as many transistors in it as those two its likely thats what they are shooting for.

 

[Orangey]

Some real fuzzy math going on itt.

 

[AllBodies]

Right so the Transistor deficit from the Fury-X is only 11%, the faster HBM2 buffer and architectural improvements should easily make that up.

I wouldn't be at all surprised if this thing is as fast as a Fury-X / GTX 980TI or perhaps a little faster. If they can pack almost as many transistors in it as those two its likely thats what they are shooting for.

Ok I feel better about this 232mm2 chip now, haha.

Also around 8 Billion transistors at 232mm2 would align with the rumours that the very largest Polaris and Pascal chips in the generation will be 18 Billion and 17 Billion respectively.

Since that would require ~520mm2 from AMD, or ~560mm2 from Nvidia.

Some real fuzzy math going on itt.

This math is perfectly legitimate

 

[bru]

So with all that, a full size 450mm^ part would be rocking at the very least 2* the performance of the 290x that's better than 295x2 performance and miles ahead of a FuryX, which, when it was talked about before, people were told they were crazy for expecting such performance. so if this is right now, then those people were correct before.

 

[AllBodies]

So with all that, a full size 450mm^ part would be rocking at the very least 2* the performance of the 290x that's better than 295x2 performance and miles ahead of a FuryX, which, when it was talked about before, people were told they were crazy for expecting such performance. so if this is right now, then those people were correct before.

As long as it scales well, yes. Sadly doubling transistors doesn't always double performance.

Though AMD have been touting the GCN revisions make it scale much better. Let's hope they aren't telling porkies.

 

[humbug]

Ok I feel better about this 232mm2 chip now, haha.

Also around 8 Billion transistors at 232mm2 would align with the rumours that the very largest Polaris and Pascal chips in the generation will be 18 Billion and 17 Billion respectively.

Since that would require ~520mm2 from AMD, or ~560mm2 from Nvidia.




This math is perfectly legitimate

A GPU with that many Transistors would be a monster, its Crysis 3 on highest settings @ 4K right there.

 

[AllBodies]

A GPU with that many Transistors would be a monster, its Crysis 3 on highest settings @ 4K right there.

Well if VR takes off (and I hope it will), we'll need to race towards being able to do 90+ FPS at 4K so we can get genuinely invisible pixels.


EDIT: For reference the 2160x1200 resolution of the first gen. Rift and Vive is only 25% more pixels than 1080p. So should be easy to do on this new gen of cards.

Whereas the 4K VR resolution I guess will be 3888x2160 due to the different aspect ratio. And that'll be 4.05x the pixels of 1080p. Might be doable on one of the largest 10nm cards.

I suppose 4320x2400 is technically '5K' on VR aspect ratio.

 

[Orangey]

Won't it be 2160x2400 x 2?

You won't need 7nm for 90fps @ 4K, just a little extra effort on the part of developers. Also bear in mind it's likely to be 2020 before a 7nm high-end GPU can even be made.

 

[AllBodies]

Won't it be 2160x2400 x 2?

You won't need 7nm for 90fps @ 4K, just a little extra effort on the part of developers. Also bear in mind it's likely to be 2020 before a 7nm high-end GPU can even be made.

That'd be 4320x2400 since you only x2 the horizontal part (since first gen is 2160x1200 or 1080x1200 per eye).

I edited my post to realise it should be 3888x2160 though. And that should be fine on 10nm.

 

[r7slayer]

4k VR? That is essentially two 4k resolutions being rendered? Jeeze most people struggle to play games on a single 4k display.


I do believe we will need at least two cards for VR when it arrives to play it optimally.

 

[drunkenmaster]

So with all that, a full size 450mm^ part would be rocking at the very least 2* the performance of the 290x that's better than 295x2 performance and miles ahead of a FuryX, which, when it was talked about before, people were told they were crazy for expecting such performance. so if this is right now, then those people were correct before.

Who said a big chip on 14nm wouldn't be significantly faster than a Fury X? No one, so congratulating yourself for being right while others were wrong about a discussion that never happened is a bit weird.

AMD aren't bringing a 450mm^2 sized core THIS YEAR, so we won't have FuryX destroying performance THIS YEAR, but when the big core comes next year they can which is the only thing people have actually been saying and yes, they are correct and you are wrong.

28nm started volume production in Q3 2011, 680gtx launched March 2012 @ 300mm^2, Titan launched 21 Feb 2013 at 561mm^2, Titan X launched March 2015 @ 600mm^2.

It was around 15 months between volume production of the process and Nvidia bringing a larger than 300mm^2 die to consumers. A further 2 years to get to a 601mm^2 part.

When do you think the last time someone made a big core early on a process was? 28nm was as above a long time after the process was available.

285gtx(simple straight optical shrink) 55nm came in Jan 2009 while AMD released their first 55nm, the 3870 in Nov 2007.

280gtx itself was June 2008 while AMD made their first 65nm parts in June 2007.

So 3870 65nm June 2007 > + 6 months > 8000gts(direct optical shrink) Dec 2007 > + 6 months > 280gtx 65nm June 2008 and Nvidia delivers a big core.

Then looking at 90nm, Nvidia had the 8800gtx on Nov 2006, AMD launched the x1900xt on 90 in Jan 2006 and the 80nm x1950xt in Oct 2006. AMD managed to get a 80nm part out before Nvidia got a 90nm big core out.

Big parts do not come early on a process. 480gtx was Nvidia's pretty much earliest attempt to bring a big core and 'match' AMD in releasing early on a process and it was a disastrous failure. It took Nvidia multiple respins and finally released a full part with all shaders working a full year after it was supposed to launch.


AMD/Nvidia can easily make things ballpark twice as fast as their 28nm parts with big cores, just big cores aren't coming this year because that is the reality of the industry. Never has a 'big' core been done within a year of a process being released(successfully, 480gtx tried, failed and still took over a year for a full/real version to launch) and this process/node is as complicated, expensive and difficult as a node has ever been.

Yeah, everyone is desperate and the transition to double patterning for sub 22nm nodes has been incredibly bad for EVERYONE, Intel having huge huge delays to 14nm as well. But it doesn't change the industry, physics or the reality of the situation. Small/medium cores early on a process node, relatively easy, big cores.... nope. Nope at 90nm, nope at 65nm, nope at 55nm, nope at 40nm and 28nm, but yes for 14/16nm? Hell no.

 

[CAT-THE-FIFTH]

Well I really do hope that isn't right or it isn't a GPU.

The card that Raja first held up to the press at CES was supposedly smaller than Cape-Verde 123mm^, so if this is the bigger Polaris and it is only 232mm^ then, that isn't going to get the job done.

There is no way that a GPU that small can be that efficient and be that powerful and if AMD think that a new top drawer chip for this year being slower than their currant top chip and having more memory is a good idea, then wow, really.

The only way that a chip that small could be faster then the FuryX is if, as Some of us said many, many months ago the performance from this new 14/16nm node is two maybe two and a half times the currant performance. When we said this we were shouted down saying how stupid and ridiculous that idea was.

I really do pray to god that this isn't the bigger of the two Polaris's that are coming sometime before September.

People forget that Hawaii has a 512 bit memory controller and transistors dedicated towards DP performance which won't do much for gaming,so the bigger chip will probably not have this,and I expect it will have a 256 bit memory controller.

I still expect it to be around GTX970/R9 290(DX11) to GTX980/R9 390X(DX12) level performance dependent on the game if it is 232MM2.

I don't understand why people are all surprised - most of the AMD midrange chips of the last few years have been 200MM2 to 250MM2,and you even have to look at the Steam Hardware Survey which is biased towards hardware enthusiasts probably answering - the vast majority are not even running a GTX970 or R9 290.

AMD has indicated this will be priced quite reasonably if the recent comments are true.

I would be quite happy with a £200 chip with decent performance.

The sub £200 area has been neglected for years and I would rather have another HD4800 series moment.

£300+ cards might be great for forum members and people running high end rigs,but its all those slower cards which are going to be the major limitation for PC games anyway.

If AMD is targetting those people in the first wave of releases and notebooks than it makes more sense. After all look at where the GM107 and GM206 have done well - lots of laptops and prebuilt desktops and Nvidia has literally swamped the market with them.

Plus,I think people were being unreasonable in thinking that the GTX980TI and Fury X which launched only in the summer would be replaced within a year?? Thats a pretty short lifespan.

I expect we will see the high end GPUs in consumer cards from both towards the end of the year,or early 2017 which would give them around 18 months of lifespan.

 

[bru]

I wasn't suggesting that the large cores would be arriving anytime soon, as has been said probably not even this year, as much as we would like it to be different. I'm just trying to say that if as has been suggested that this small chip is going to be up there with the FuryX, then the big chip to follow will be as quick as was suggested previously.