AMD 20nm GPU Discussion Thread

[Techhog]

AMD has played their GPU cards close to the chest since the 4870 days, which has caught Nvidia off guard almost every time. They won't show off any 20nm press stuff until its so close to launching that the only thing Nvidia could do to respond would be price cuts.

I really highly doubt that there's zero collusion in a duopoly like this.

 

[el etro]

Well, lets begin the speculation(again):


Pirate Islands 1:

-Same 28nm process;
-50% more perf/watt over last architecture in overall.
-Performance bets:
*Mid-end chip(HD7870 size): R9 280x performance;
*High-end chip(HD7970 size): R9 290x(quiet) performance;
*Top-end chip(290x size?): 10% faster than 7990.
*Halo class dual GPU card: equals 290x CFX performance.
-End of 2014 launch.



Pirate islands 2:

-16nm process(FinFet):
-150% more perf/watt(arch+manufacturing tech improvements) over 28nm cards;
-Performance bets:
*Mid-end chip: R9 290 performance;
*High-end chip: GTX 690 performance;
*Top-end chip: Ties R9 290(Non-x) crossfire/GTX 780(Non-Ti) Sli;
*Halo class dual GPU card: Better than 290x2 performance.
-Xmas 2015 launches 1st card.

 

[DiogoDX]

1H 2014: AMD possibly tapes out 20nm GPU / SK Hynix gives HBM samples to partners *cough*AMDisone*cough*.

2H 2014: AMD starts 20nm GPU production / SK Hynix starts HBM production.

Coincidence?! I think not! Let the conversation BEGIN!

AMD (ATI precisely) is the creator of GDDR3, GDDR4 and GDDR5 standards so would not be surprise that they came first with the stack dram.

 

[ShintaiDK]

AMD (ATI precisely) is the creator of GDDR3, GDDR4 and GDDR5 standards so would not be surprise that they came first with the stack dram.

There are multiple versions of stacked DRAM. And no, AMD will not be first. Knights Landing will using Microns version. And by the looks of it, nVidia will be next.

You make it sound like AMD was the sole creator of GDDR3/4/5, that is wrong as well.

 

[DiogoDX]

There are multiple versions of stacked DRAM. And no, AMD will not be first. Knights Landing will using Microns version. And by the looks of it, nVidia will be next.

You make it sound like AMD was the sole creator of GDDR3/4/5, that is wrong as well.

Not sole but a important part. I always like to reference this article that is a good read about the topic: http://theovalich.wordpress.com/200...r5-analysis-or-why-gddr5-will-rule-the-world/

 

[ShintaiDK]

Not sole but a important part. I always like to reference this article that is a good read about the topic: http://theovalich.wordpress.com/200...r5-analysis-or-why-gddr5-will-rule-the-world/

Written by one of the most untrusted people 😉

 

[ShintaiDK]

-16nm process(FinFet):-Xmas 2015 launches 1st card.

These 2 points are unlikely to match.

Also its not 16nm, but 20nm with Finfets. And 2015 is _very_ optimistic for that. If you mean the real 16nm, then its even further away.

I doubt AMD and nVidia can afford to skip the 20nm non finfet node. Because that would mean 5 year+ with 28nm. Their business model simply cant handle that.

 

[Gloomy]

I doubt AMD and nVidia can afford to skip the 20nm non finfet node. Because that would mean 5 year+ with 28nm. Their business model simply cant handle that.

Why not? Who is competing with them?

 

[ShintaiDK]

Why not? Who is competing with them?

Their own cashflows.

 

[Gloomy]

Their own cashflows.

Surely it would be cheaper to stay on this node, and design more chips in it.

 

[NTMBK]

These 2 points are unlikely to match.

Also its not 16nm, but 20nm with Finfets. And 2015 is _very_ optimistic for that. If you mean the real 16nm, then its even further away.

I doubt AMD and nVidia can afford to skip the 20nm non finfet node. Because that would mean 5 year+ with 28nm. Their business model simply cant handle that.

The "16nm" node does see some chip size improvements, but obviously not as much as a traditional node shrinks:

In fact, Dr Liu claims TSMC has the better performance. "We took the approach of using the FinFET transistor to improve the performance on top of the similar back end technology in our 20nm process. This transistor performance, and innovative layout methodology, can improve the chip size by about 15%. Because transistor drive is much stronger, you don't need such a big area to deliver the same driving circuitry."

http://www.newelectronics.co.uk/ele...ransistor-density/59122/#sthash.yzUx2Fvk.dpuf

And the FinFET node is meant to be coming very rapidly after the 20nm node, as in within a year. Given that NVidia is rolling out a whole new line of 28nm Maxwell GPUs, I suspect that they are digging in to wait out the period until FinFETs are available. The roadmap says Pascal in 2016; that says to me that they are waiting for FinFETs.

Don't forget that the 20nm node is "SoC", i.e. low power- the high performance process got canned a while back. It's not beyond the realms of possibility that a shrink of a 28nm GPU to 20nm would give worse performance. If NVidia was going to use 20nm anywhere, it would use it on its low power mobile chips- but we just saw them bringing out a brand new GM107 on 28nm. If they want to recoup the design costs then they're going to have to not replace it for quite a while yet.

 

[NTMBK]

Why not? Who is competing with them?

Intel. If Intel keeps tick-tocking their graphics into more and more products, the market for discrete GPUs is in trouble. NVidia is already getting squeezed out of the Macbook Pro and iMac lineup.

 

[ShintaiDK]

Surely it would be cheaper to stay on this node, and design more chips in it.

Sales would drop. Meaning smaller cashflow. And there is only so much you can do on the same node. AMD got GCN 1.1 as their new uarch. nVidia got Maxwell. GCN 1.1 barely offers anything. And Maxwell on 28nm while great performance/watt wise, is still not what a 20nm chip would bring. And people are not upgrading their old cards for that.

And Broadwell with expanded GT3e models on 14nm is also eating its share. Mobile is already a lost cause fo dGPUs.

 

[ShintaiDK]

The "16nm" node does see some chip size improvements, but obviously not as much as a traditional node shrinks:



http://www.newelectronics.co.uk/ele...ransistor-density/59122/#sthash.yzUx2Fvk.dpuf

And the FinFET node is meant to be coming very rapidly after the 20nm node, as in within a year. Given that NVidia is rolling out a whole new line of 28nm Maxwell GPUs, I suspect that they are digging in to wait out the period until FinFETs are available. The roadmap says Pascal in 2016; that says to me that they are waiting for FinFETs.

Don't forget that the 20nm node is "SoC", i.e. low power- the high performance process got canned a while back. It's not beyond the realms of possibility that a shrink of a 28nm GPU to 20nm would give worse performance. If NVidia was going to use 20nm anywhere, it would use it on its low power mobile chips- but we just saw them bringing out a brand new GM107 on 28nm. If they want to recoup the design costs then they're going to have to not replace it for quite a while yet.

By whole line of 28nm Maxwell GPUs. You mean GM107 and GM108? The model numbers for the others are GM204, GM206 and GM200 and verified by drivers.

And that SOC process isnt as bad as its portrait. My i5 4670 is essentially build on the same.

TSMC and statements are something that doesnt go hand in hand. So lets see when they actually delivers. We all know how incorrect TSMCs own roadmaps are.

Late 2014 seems be a battle between 20nm Maxwell and 20nm GCN 1.1.

 

[sontin]

And Broadwell with expanded GT3e models on 14nm is also eating its share. Mobile is already a lost cause fo dGPUs.

Nonsense:

http://www.anandtech.com/show/7834/nvidia-geforce-800m-lineup-battery-boost/4

 

[ShintaiDK]

Nonsense:

http://www.anandtech.com/show/7834/nvidia-geforce-800m-lineup-battery-boost/4

Yes, the slide is PR nonsense.

 

[jpiniero]

If anything, OEMs have pretty much ignored GT3e, except for Apple. Broadwell ULT (IIRC) doesn't support discrete graphics, so there's definitely some risk to both AMD and nVidia though.

 

[NTMBK]

By whole line of 28nm Maxwell GPUs. You mean GM107 and GM108? The model numbers for the others are GM204, GM206 and GM200 and verified by drivers.

Yes, all of those Maxwells. I flat out don't believe that NVidia would create a 28nm Maxwell design solely to produce GM107 and GM108. These are chips destined for the laptop market, where power savings would be utterly vital, and yet these are the only ones in the whole lineup built on 28nm? It just doesn't add up. In your scenario the only reasonable explanation for them going 28nm is time-to-market- but NVidia already has the laptop dGPU market stitched up, no need to rush out a 28nm Maxwell just for that. It could be a pre-emptive reaction against Broadwell, I guess, but Broadwell is not hitting until Q4- roughly the same time as the other Maxwells are meant to arrive.

And that SOC process isnt as bad as its portrait. My i5 4670 is essentially build on the same.

No, your 4670 is build on Intel's higher performance variant of its 22nm process. Its true SoCs (Bay Trail, Merrifield) are built on the lower power variant of the 22nm node. That is the process which is equivalent to 20nm SoC (in terms of power envelope, not tech, obviously).

TSMC and statements are something that doesnt go hand in hand. So lets see when they actually delivers. We all know how incorrect TSMCs own roadmaps are.

Hah, yeah, I got a good laugh looking at some old roadmaps saying that 20nm was coming in 2013 and 16nm in 2014. But despite the roadmaps frequently slipping, they have consistently had the gap between 20nm and 16nm at ~1 year. Given how similar the two processes are, it's not hard to believe that the gap would not be very large. Not the mention the canning of the 20nm HP process would free up resources to work on 16nm, and provide a very good incentive to get 16nm out the door.

Late 2014 seems be a battle between 20nm Maxwell and 20nm GCN 1.1.

Apart from the 20nm part, I agree. 😉

 

[ShintaiDK]

In terms of the 20nm SOC.

TSMC only got 1 processnode for all its future. Do you think it will be anything else?

Its all SOC nodes. 20nm without FF, 20nm with FF, 16nm with FF.

 

[el etro]

Also its not 16nm, but 20nm with Finfets. And 2015 is _very_ optimistic for that.

If you mean the real 16nm, then its even further away.

I know its not real 16nm. It was scheduled to tape-out in 2014.

 

[ShintaiDK]

I know its not real 16nm. It was scheduled to tape-out in 2014.

And 20nm taped out in 2011 arcording to GloFo:
http://www.eetimes.com/document.asp?doc_id=1260807

So its amazing the same 20nm node took so long to tapeout again 😉

They just barely started shipping 28nm in 2014. Roadmap said 2011.

 

[el etro]

Nonsense:

http://www.anandtech.com/show/7834/nvidia-geforce-800m-lineup-battery-boost/4

I can't believe in total system power showed in this graph, even if 840m is a Maxwell chip.

 

[raghu78]

AMD 20nm GPUs will sport an improved GCN 2.0 architecture and most likely HBM (stacked memory).

http://electroiq.com/blog/2013/12/amd-and-hynix-announce-joint-development-of-hbm-memory-stacks/

I am guessing 4096 stream processors @850 - 900 Mhz (because TSMC 20nm will be a power constrained node), 64 CU, 64 ROPs and 4096 bit memory (4 HBM stacks running at 1 Ghz). With architectural improvements and improved per core performance a 40% improvement over R9 290X is possible.

http://sites.amd.com/us/Documents/TFE2011_006HYN.pdf

 

[chimaxi83]

Damn, I came in here to read about AMD 20nm info, only to find a boatload of Nvidia stuff. Weird.

 

[Aeiou]

I just hope the high end lets me max The Witcher 3 at an insane res, that's all i care about.