Polaris foundry, the official word

[JDG1980]

If the Polaris demo chip was indeed fabbed on GloFo 14nm (rather than Samsung) then that's very good news. Up until now we haven't had any reason to believe GloFo had actually produced anything except PowerPoint slides on 14nm. If they've got working products taped out then that gives some hope that we won't see major delays to mass production.

AMD needs to beat Nvidia to market, and a chip that fills the GM107 niche (mobile and <75W desktops) would be an excellent choice; it has the potential to be a big seller. Apple, especially, would be happy to see this so that they can give the big MacBook Pro something better than 2012-vintage Cape Verde.

 

[PPB]

Prepare to be surprised. I mean, Anandtech claims they will, while the question the OP fielded specifically mentioned 16nm and the implied answer is yes. Most convincing is the fact that AMD is calling the process simply "FetFET." If they were only using 14nm, they'd call it 14nm. Especially given that Nvidia's entire stack is "only" on 16nm.

Not calling it specifically by the 14nm can be dubious. Then again, that node is only 14nm in its name and the relevant part of it (using Finfets) is indeed named.

He didnt just reply "yes". And in his reply he didnt mention nodes, but foundries. Little difference in wording might end up in a big difference in reality vs people's expectations.

At this point I have very little confidence in GloFlo or TSMC. However, we know that the former is currently producing working 14nm GPUs.

I think I remember reading that both Polaris chips were on 14nm. If that's true, then given that AMD is also producing some chips on TSMC 16nm that haven't been shown off and Nvidia can't seem to get back anything demo-able from TSMC it appears GloFo is way ahead on FinFet GPUs.

Now this is FUD material.

You cant even know they are even using 2 different nodes and are ready to jump to the conclusion that Glofo is ahead of TSMC (????????)

 

[Elixer]

Lots of discussion here on where Polaris is being built. Some think it will be GloFo exclusive, others think it will be GloFo/TSMC split.

I contacted AMD and was told that AMD is "engaging with multiple foundries" for Polaris and that the part they demo'd (the GTX 950 competitor) was built on GloFo 14nm.

This isn't anything new, they have been saying that for a long time.
What would be new is if they mention who exactly they are using, and that data will only show on the quarterly reports.

 

[Silverforce11]

Anandtech still thinks GCN 1.1+ can only do 8 compute queues in parallel 🙄 Talk about quoting a tech site as the absolute truth.

Just google "confirmation bias", take a nap and think why you are so obsessed with Polaris being 2 different nodes.

And Anandtech still thinks Maxwell is capable of Async Compute. lol

They refuse to fix a blatantly wrong article when they were lied to by NV PR.

AMD themselves have said publicly GCN is the only uarch that can do Async Compute! Still, no corrections to their article...

https://community.amd.com/community...blade-stealth-plus-razer-core-gaming-solution

AMD's advanced Graphics Core Next (GCN) architecture, which is currently the only architecture in the world that supports DX12's asynchronous shading.

Regarding the node for Polaris, this is confirmation already: http://english.etnews.com/20151222200002 along with Raja's interview where he specifically mentions Polaris uarch is designed for 14nm ff.

According to industries on the 21st, Samsung Electronics&#8217; System LSI Foundry Business Department will start mass-producing AMD&#8217;s new GPU &#8216;Greenland (development code name)&#8217; along with Global Foundry (GF) starting from 2nd quarter of 2016. Greenland will be produced from Gen. 2 14-nano FinFET LPP (Low Power Plus) processing and its electricity efficiency per watt is 2 times higher compared to 28-nano GPU (code name: Fiji) that is currently being sold in markets. Proportion of production supplies is very fluid as AMD will regulate proportion between 2 businesses according to many conditions such as yield and others.
AMD had been asking TSMC from Taiwan to be in charge of GPU production but it decided to cut contract with TSMC with 28-nano being the last production after issues with yield and instable supplies had continued.

This is the important part:

&#8220;Because Samsung Electronics and GF have same IP for 14-nano processing, chips that are designed by AMD will all be produced at both factories.&#8221; said a person who is familiar with this industry. &#8220;If products are produced from both factories, AMD won&#8217;t have to worry about a problem regarding lack of supplies.&#8221;
After starting production of Greenland, AMD will put out its new CPU &#8216;Zen&#8217; right after. This product is also produced from 14-nano LPP processing and will be produced by Samsung Electronics and GF just like GPU.

Do you people think AMD has the R&D $$ to design for 2 separate and different nodes?? After all the talk about how AMD is financially crippled, investing less and less in R&D, it makes sense for them to pour heaps of $$ in 2 different designs? Recipe for failure right there.

Samsung Electronics&#8217; Foundry Business Department is very encouraged after it took over customers from TSMC, which is Samsung Electronics&#8217; biggest competitor, one after the other. It is very meaningful that Samsung Electronics was able to expand its foundry products that were limited to mobile products to GPU and CPU for PCs.

If GloFo can't execute with good volume, Samsung will. Eggs in two basket, with one design. The logical choice.

 

[HurleyBird]

You cant even know they are even using 2 different nodes and are ready to jump to the conclusion that Glofo is ahead of TSMC (????????)

Hence "if true" and "appears." All evidence so far points to Glofo being ahead, but it's not particularly strong evidence.

 

[Stuka87]

Prepare to be surprised. I mean, Anandtech claims they will, while the question the OP fielded specifically mentioned 16nm and the implied answer is yes. Most convincing is the fact that AMD is calling the process simply "FetFET." If they were only using 14nm, they'd call it 14nm. Especially given that Nvidia's entire stack is "only" on 16nm.

Just because AT says it, doesn't mean its true. I highly doubt AMD would design chips for two different technologies. GF and Samsung use the same process, so it makes complete sense that they would use those two, and not TSMC.

 

[Vesku]

I don't think AMD wants to cut their relationship with TSMC completely, that would put them in a terrible position if TSMC is first to the next node shrink. Whether that will mean a Polaris dGPU at TSMC or some console/embedded SoCs remains to be seen. If 16FF+ has good yields with large dies I could see AMD making the largest HPC/Enthusias dGPU at TSMC.

 

[gamervivek]

Just because AT were wrong about async compute queues and stuff doesn't mean that they are not well reliable in this case.

AMD will most likely get bigger chips from TSMC's 16nm and the smaller ones from the 14nm. The supposedly Polaris chips going around on zauba are much less pricier than the Fury X chips. So they're more likely cape verde and pitcairn in size, the latter being around Hawaii in performance.

 

[Azix]

Just because AT were wrong about async compute queues and stuff doesn't mean that they are not well reliable in this case.

AMD will most likely get bigger chips from TSMC's 16nm and the smaller ones from the 14nm. The supposedly Polaris chips going around on zauba are much less pricier than the Fury X chips. So they're more likely cape verde and pitcairn in size, the latter being around Hawaii in performance.

but where does this assumption that TSMC would be better for big chips come from?

 

[Glo.]

but where does this assumption that TSMC would be better for big chips come from?

Because of A9 produced both on TSMC and Samsung. People believe it because, in mind, TSMC has better process, thats why they owned the deal with Apple.

The truth is that Apple had to buy capacity from TSMC, and A9X was not enough. Of course, performance of A9X would not be achievable on 14 LPE, but still, the reason for producing A9 on TSMC fabs was only the capacity that Apple had to buy from TSMC.

Mindshare here plays the biggest role, not really the reason.

 

[Dribble]

but where does this assumption that TSMC would be better for big chips come from?

Previous experience - TSMC have been making monster high power gpu chips for many years. What's the biggest chip that Samsung fab? - the low powered mobile soc's they are famous for are very different from a high end gpu.

 

[MrTeal]

Previous experience - TSMC have been making monster high power gpu chips for many years. What's the biggest chip that Samsung fab? - the low powered mobile soc's they are famous for are very different from a high end gpu.

GloFo has made some large dies in the past, and their recently acquired East Fishkill plant has put out dies (and continues to do so) that are larger than even GM200. It's not their 14nm process, but then TSMC hasn't made 600mm 14nm dies either.

 

[alcoholbob]

Shrinking an architecture from 28nm to 16nm should be no easier than porting a 14nm architecture to 16nm...

Not to mention what they are calling 16nm FF is just 20nm silicon with 16nm FinFet gates, and theyve been preparing to transition to 20nm since 2013.

 

[Arachnotronic]

but where does this assumption that TSMC would be better for big chips come from?

TSMC's yields are SIGNIFICANTLY better than GloFo's yields on Samsung's 14nm process. Samsung's 14nm yields are already lower than TSMC's 16nm last I asked around and GloFo's implementation of Samsung's 14nm is far, far worse yielding.

AMD has a WSA to satisfy so they'll do the small chips on GloFo 14nm, but where they need the performance and yields to fight the big Pascal guns, they will use TSMC.

Anybody who "wishes" for AMD to use GloFo 14nm is just hoping for AMD to enter this fight with one arm tied behind its back. The only way AMD can be process competitive with the big Pascals is to use the best process available -- TSMC 16FF+.

 

[Glo.]

Time will only show where truth is.

 

[PPB]

TSMC's yields are SIGNIFICANTLY better than GloFo's yields on Samsung's 14nm process. Samsung's 14nm yields are already lower than TSMC's 16nm last I asked around and GloFo's implementation of Samsung's 14nm is far, far worse yielding.

AMD has a WSA to satisfy so they'll do the small chips on GloFo 14nm, but where they need the performance and yields to fight the big Pascal guns, they will use TSMC.

Anybody who "wishes" for AMD to use GloFo 14nm is just hoping for AMD to enter this fight with one arm tied behind its back. The only way AMD can be process competitive with the big Pascals is to use the best process available -- TSMC 16FF+.

Prove with official information regarding yields on both process nodes from those 2 companies (if there is any publicly available), then try to make an argument. You are already making 5 BS posts in a row with no real information whatsoever.

Some people just need to start writing fiction. At least they would get some money out of their hobby.

 

[Arachnotronic]

Prove with official information regarding yields on both process nodes from those 2 companies (if there is any publicly available), then try to make an argument. You are already making 5 BS posts in a row with no real information whatsoever.

Some people just need to start writing fiction. At least they would get some money out of their hobby.

We're only allowed to talk about official information that comes from PR departments and not from actually doing the legwork required to figure out what's going on in the industry?

Well then, better put me on ignore :thumbsup:

 

[tential]

I think it's great you asked. Think you can follow up ask and see if any gpu will be made at tsmc this next round? I think that's an answer they can confirm

 

[USER8000]

Isn't Samsung/Globalfoundries being used for CPU products like Zen too?? Weird an older process which has been in large scale production for quite a while for a large volume production product,aka,The Samsung S6 which was launched last year in April and uses a 14NM SOC,seems to be having massive yield issues.

 

[AnandThenMan]

We're only allowed to talk about official information that comes from PR departments and not from actually doing the legwork required to figure out what's going on in the industry?

You said:

TSMC's yields are SIGNIFICANTLY better than GloFo's yields on Samsung's 14nm process.

You were asked for a source, so please provide or if you are just making it up then say so.

 

[3DVagabond]

TSMC's yields are SIGNIFICANTLY better than GloFo's yields on Samsung's 14nm process. Samsung's 14nm yields are already lower than TSMC's 16nm last I asked around and GloFo's implementation of Samsung's 14nm is far, far worse yielding.

AMD has a WSA to satisfy so they'll do the small chips on GloFo 14nm, but where they need the performance and yields to fight the big Pascal guns, they will use TSMC.

Anybody who "wishes" for AMD to use GloFo 14nm is just hoping for AMD to enter this fight with one arm tied behind its back. The only way AMD can be process competitive with the big Pascals is to use the best process available -- TSMC 16FF+.

FWIU this is a different "gen 2" process for Polaris, LPP. Not what they've been using up until now, LPE.

 

[beginner99]

Do you people think AMD has the R&D $$ to design for 2 separate and different nodes?? After all the talk about how AMD is financially crippled, investing less and less in R&D, it makes sense for them to pour heaps of $$ in 2 different designs? Recipe for failure right there.

If GloFo can't execute with good volume, Samsung will. Eggs in two basket, with one design. The logical choice.

Exactly. They had to design for a new process and going 14 nm exclusively just makes so much sense for AMD. They can reuse knowledge of the process for both GPUs and for Zen. Anything else woudl again be a stupid move from AMD.

 

[beginner99]

TSMC's yields are SIGNIFICANTLY better than GloFo's yields on Samsung's 14nm process. Samsung's 14nm yields are already lower than TSMC's 16nm last I asked around and GloFo's implementation of Samsung's 14nm is far, far worse yielding.

So? If they get 14 nm cheaper than 16 nm and every defect die still makes AMD money because less to pay later due to WSA. Bad yield also doesn't mean bad performance. The good dies might actually perform better than anything 16nm. We don't now but it's possible.

I just don't see how AMD could go for 2 nodes. That would cost them at least double-digit millions more than sticking with 1 node.

 

[AtenRa]

My prediction,

Polaris 10 and 11 is strictly 14nm, no need to spend money and time to design and make new masks for two different processes for only two dies. Not even Intel would do that.

But, AMD may use TSMC 16nm FF+ for its new semi-custom deal starting from H2 2016. They could also use TSMCs 16nm FF+ in 2017 onward for a third or even forth GPU die.

 

[guskline]

Since AMD already has Polaris silicon and previewed it, am I to assume it is the Polaris 10? Didn't Raji say Polaris 11 would be the high end?

Which Polaris was shown to a group but pictures were not allowed to be taken and some details are still under wraps?