[ WCCF Tech ] TSMC will produce 16nm FinFET Nvidia Geforce GPUs – No Word For AMD

[KaRLiToS]

SOURCE

We covered TSMC’s plans to produce the next generation of Nvidia’s GPUs a while back and now an even more solid report has been released from DigiTimes. TSMC is getting ready to manufacture Nvidia’s 16nm FinFET GPUs and there is a very high chance now that we see 20nm process skipped entirely. However, something that is very interesting is the fact that AMD’s name is once again missing from the present partners list.

However, there is just a sliver of uncertainty left as this is still a report from third party sources. TSMC will be holding its annual suppliers conference on Thursday and we will know for certain whether AMD’s name was just further down the list or completely absent as we suspect. The details of 16nm and 10nm processes will be shared and all the major semiconductor partners will be there including Nvidia.

 

[Silverforce11]

Hmm, maybe GloFo finally picked up their game?

 

[tviceman]

The competition among the foundries seems to be heating up. Samsung and GloFo have teamed up in their R&D research; perhaps GloFo got a shot in the arm and/or is cutting AMD a deal.

I like the idea of 20nm GPU's being skipped entirely. I think Tegra 6 aka Erista will be on 20nm though, but that is an entirely different subject.

 

[alcoholbob]

What's the rush with being named on the 16FF list if TSMC isn't going to be in mass production of any of these until 2016 anyway.

 

[IEC]

Are they going straight to 16nm FinFET because 20nm is planar?

Apparently:
https://www.semiwiki.com/forum/content/1789-16nm-finfet-versus-20nm-planar.html

 

[96Firebird]

perhaps GloFo got a shot in the arm and/or is cutting AMD a deal.

I hope so, ever since Fab 8 opened up here in upstate NY, I've been rooting for them to get some wins. I even looked at a couple engineering jobs there when I graduated, but they all had crazy hours that I wasn't prepared to work. Submitted to a couple anyways, just to get a free tour. Never happened though...

Hope to see some US fabs competing with Taiwan.

 

[nenforcer]

Are they going straight to 16nm FinFET because 20nm is planar?

Apparently:
https://www.semiwiki.com/forum/content/1789-16nm-finfet-versus-20nm-planar.html

Global Foundries is partnering with Samsung for 14nm FinFET

http://www.globalfoundries.com/news...ering-of-14nm-finfet-semiconductor-technology

AMD and Global Foundries are pretty tight so I would expect AMD's next generation CPU and GPU to be on this process done by Global Foundries in 2016.

I think AMD Radeon R9 390/X due out early next year is still only .28nm so this could be done at either TSMC or GlobalFoundries?

 

[MeldarthX]

Pretty sure also due to the new agreement AMD has to bring everything back to Gloflow; that includes gpus.

We're starting to see some of the goodness out of gloflo with other parts......

 

[Bateluer]

I think AMD Radeon R9 390/X due out early next year is still only .28nm so this could be done at either TSMC or GlobalFoundries?

390X will likely be 20nm, not 28nm, based on AMD's public statements.

AMD's CEO has said they will have 20nm parts in early 2015.

“20nm is an important node for us. We will be shipping products in 20nm next year and as we move forward […],” said Lisa Su, senior vice president and chief operating officer of AMD. “If you look at our business, it is quite a bit more balanced between the semi-custom, embedded, […] professional graphics […] as well as the more traditional sort of client and graphics pieces of our business. [20nm] technology plays in all of those businesses.”

http://wccftech.com/amd-20nm-gpus-horizon-tsmc-ramps-20nm-production/

 

[NTMBK]

390X will likely be 20nm, not 28nm, based on AMD's public statements.

AMD's CEO has said they will have 20nm parts in early 2015.



http://wccftech.com/amd-20nm-gpus-horizon-tsmc-ramps-20nm-production/

lol, a WCCFtech story which cites an Anton Shilov story which misquotes Lisa Su and doesn't cite anything. Fantastic.

Here is the actual quote direct from the transcript (with added emphasis by me):

Sure, John. Let me take a stab at that. I think when you look at what's important to us, clearly process technology's an important element. But, we have invested quite a bit in architecture, design techniques, new IP software.

So, I wouldn't say that process technology is the first and primary indeterminate for us. It is important that we are on competitive technology.

So, we said before and I will say again, that 20-nanometer is an important node for us. We will be shipping products in 20-nanometer next year. And as we move forward, obviously, FinFET is also important.

If you look at our business, it is quite a bit more balanced between the semi-custom, embedded, commercial pro-graphics growth portions as well as the more traditional client and graphics pieces of our business. Technology plays in all of those businesses.

Transcript available at http://phx.corporate-ir.net/External.File?item=UGFyZW50SUQ9MjQyNDQ3fENoaWxkSUQ9LTF8VHlwZT0z&t=1

She is dodging the question, and openly saying that process tech is not the only way that they will advance their products in 2015. She says that they will launch 20nm products, but that is not news- we already know about the ARM and x86 Skybridge APUs, which are on 20nm. She is not committing to 20nm discrete GPUs, just to new "technology".

This is very clear to anyone who reads the original source, but of course Anton selectively misquotes Lisa Su so that he can get his sensationalist headline- because that is how he gets clicks, and ad revenue.

My bet is 28nm GPU with stacked memory.

 

[MeldarthX]

lol, a WCCFtech story which cites an Anton Shilov story which misquotes Lisa Su and doesn't cite anything. Fantastic.

Here is the actual quote direct from the transcript (with added emphasis by me):


Transcript available at http://phx.corporate-ir.net/External.File?item=UGFyZW50SUQ9MjQyNDQ3fENoaWxkSUQ9LTF8VHlwZT0z&t=1

She is dodging the question, and openly saying that process tech is not the only way that they will advance their products in 2015. She says that they will launch 20nm products, but that is not news- we already know about the ARM and x86 Skybridge APUs, which are on 20nm. She is not committing to 20nm discrete GPUs, just to new "technology".

This is very clear to anyone who reads the original source, but of course Anton selectively misquotes Lisa Su so that he can get his sensationalist headline- because that is how he gets clicks, and ad revenue.

My bet is 28nm GPU with stacked memory.

You know the 380X very well could be that and the 390X......could also be 20mn with stacked memory......so we get both......but that's just a crazy thought

 

[raghu78]

lol, a WCCFtech story which cites an Anton Shilov story which misquotes Lisa Su and doesn't cite anything. Fantastic.

Here is the actual quote direct from the transcript (with added emphasis by me):


Transcript available at http://phx.corporate-ir.net/External.File?item=UGFyZW50SUQ9MjQyNDQ3fENoaWxkSUQ9LTF8VHlwZT0z&t=1

She is dodging the question, and openly saying that process tech is not the only way that they will advance their products in 2015. She says that they will launch 20nm products, but that is not news- we already know about the ARM and x86 Skybridge APUs, which are on 20nm. She is not committing to 20nm discrete GPUs, just to new "technology".

This is very clear to anyone who reads the original source, but of course Anton selectively misquotes Lisa Su so that he can get his sensationalist headline- because that is how he gets clicks, and ad revenue.

My bet is 28nm GPU with stacked memory.

yeah. I am quite sure that AMD has gone with GF 28SHP and 2.5D stacking on silicon interposer with HBM.

http://seekingalpha.com/article/272...redit-suisse-technology-conference-transcript

Devinder Kumar - SVP and CFO

"I can tell you that with the changes that have occurred on GlobalFoundries with the management team and the focus that Abu Dhabi has and the investment that they have as a partnership with GlobalFoundries, the relationship between AMD and GlobalFoundries is the best in the history of the relationship. The folks that we're working on in fact we just had a meeting in Abu Dhabi just a couple of weeks ago. Really good discussions, very business oriented.

And I think the execution of GlobalFoundries has improved significantly and that helps us from an overall standpoint. In 2014 for the first time, some folks may not know then. For the first time in the history of the relationship we went beyond PC product and actually we are making graphics, PC, and semi-custom products at GlobalFoundries in 2014 and that continue into 2015. When you diversify the product that you make at a foundry like GlobalFoundries, it benefits them from a mix standpoint and benefits us from a mix standpoint. And like I said, the execution is continuing to get better and we are very pleased - very, very pleased with that relationship."

AMD is already manufacturing Kaveri, Beema, Mullins, game console chips and graphics at GF 28SHP. For 2015 I expect AMD to go top to bottom with GF 28SHP - GPUs, APUs (desktop,notebook,embedded), semi-custom game console chips, Seattle server SoC and any other newer server chips if planned.

Given the fact that GF has licensed Samsung 14nm FINFET and will be ramping production in H1 2015 (Samsung expects to start 14nm FINFET volume production by end of 2014.) I wouldn't be surprised if AMD is one of the first adopters of Samsung/GF 14nm FINFET. We can expect 14nm FINFET GPUs from AMD by late 2015 or early 2016.

 

[Silverforce11]

390X will likely be 20nm, not 28nm, based on AMD's public statements.

AMD's CEO has said they will have 20nm parts in early 2015.


http://wccftech.com/amd-20nm-gpus-horizon-tsmc-ramps-20nm-production/

That's CEO speak when they deliberately be vague. Yes, they are launching 20nm products. It's their APUs & SOCs.

I still believe their GPUs will be 28nm, just as NV is sticking with 28nm until TSMC 16nmFF are ready.

With GloFo having 14nmFF, the competition with TSMC can only be GOOD!

 

[exar333]

Somewhat scary for AMD honestly. GF doesn't have a great track-record, but hopefully they can improve.

 

[Keysplayr]

Somewhat scary for AMD honestly. GF doesn't have a great track-record, but hopefully they can improve.

The only thing that lends hope is that Samsung is involved. Otherwise I'd obviously be as skeptical about global foundries track record as you are. I think there is hope now.

 

[RussianSensation]

Somewhat scary for AMD honestly. GF doesn't have a great track-record, but hopefully they can improve.

Ya, seriously. My friend worked in the higher management at GloFo and directly interfaced with the CEO. He quit in September of this year and said GloFo is a total mess at a the moment with poor execution and inability to recoup fab costs. He wouldn't go into any details regarding any wins for AMD as that's confidential information but based on my discussions with him GloFo has not been able to meet internal deadlines at all. That's not great news for AMD. Secondly, we don't even know if GloFo's node is at least as good as TSMC's, when it could be far worse from a technical standpoint.

AMD could not only end up with inferior node characteristics but be way behind NV+TSMC for next gen cards. NV is well on track to deliver 16nm FinFET Pascal chips in 2016 which leaves no breathing room for AMD as they haven't even launched R9 300 series. Based on the success of GM204, being late is very costly for AMD.

 

[raghu78]

Ya, seriously. My friend worked in the higher management at GloFo and directly interfaced with the CEO. He quit in September of this year and said GloFo is a total mess at a the moment with poor execution and inability to recoup fab costs. He wouldn't go into any details regarding any wins for AMD as that's confidential information but based on my discussions with him GloFo has not been able to meet internal deadlines at all. That's not great news for AMD. Secondly, we don't even know if GloFo's node is at least as good as TSMC's, when it could be far worse from a technical standpoint.

AMD definitely knows whats the situation at the 3 foundries. If Samsung has problem with FINFET yields as few people are saying then GF is in the same boat.

http://www.eetimes.com/document.asp?doc_id=1324211
http://www.eetimes.com/document.asp?doc_id=1324617

AMD will definitely watch the developments at TSMC and Samsung / GF closely before deciding on their 2016 product stack and much before signing their 2016 WSA agreement. I am sure AMD will be making their foundry decisions soon for 2016 APUs,GPUs and CPUs. That might be the reason their Financial Analyst day got pushed to Q1 2015.

http://www.kitguru.net/components/g...nancial-day-event-as-it-rethinks-its-roadmap/
http://hexus.net/tech/news/industry/77721-amd-delays-financial-analyst-day-rethink-product-roadmap/

AMD could not only end up with inferior node characteristics but be way behind NV+TSMC for next gen cards. NV is well on track to deliver 16nm FinFET Pascal chips in 2016 which leaves no breathing room for AMD as they haven't even launched R9 300 series. Based on the success of GM204, being late is very costly for AMD.

Wait till GM200 and R9 390X are launched. What would you be saying if AMD recovers the lost market share and gains some more at Nvidia's expense in 2015. Nvidia's FINFET products are not going to be out before early 2016. btw I don't expect Pascal to launch until late 2016. I don't see Nvidia going for a new FINFET node with associated yield challenges, a new architecture, a new memory standard, a new chip production technology (2.5D stacking on silicon interposer) all in one go. It might just be too much to ask. If it happens Nvidia will surprise even themselves. I can see a Maxwell GM304 at 16FF+ in early 2016 followed by Pascal GM400 by late 2016 with all the bells and whistles. Pascal,DX12,NVLink, 2.5D stacking, 8 or 16GB HBM with 1TB/s bandwidth. That chip could be a true beast. :biggrin:

btw AMD can always fab at TSMC 16FF+ as they have done in the past. Why do you think AMD would risk falling behind badly against Nvidia ? You always seem to have a bad prediction waiting for AMD. :biggrin:

 

[SlickR12345]

AMD next GPU's are going to be 20nm, that is why they haven't released a 390x on 28nm.

We are going to see 20nm GPU's in March/April from AMD.

 

[nine9s]

Anyone know about reports that these foundries are having a difficult time at these sizes ( below 20) with extremely low yield potential making it cost prohibited to commercial make below 20nm chips for the foreseeable future (2 years or so.) I have been reading, in investment reports, the last few years that yield rates will currently be so low on these sizes at most foundries for hire that it will result in delay after delay.

This is the reason Intel has been so aggressive in foundry buildout the last few years (doubling or tripling its entire foundry capacity) - it was preparing for something big. The jest is that for commercially successful yields over the next few years, most all business in these chips sizes will have to go through Intel with Intel's only question being whether or not it will become a foundry for hire allowing ARM to survive by building chips under that architecture at those syb 20 sizes or if it will still insists on Intel design only, therefore basically killing ARM as far as new chip wins/designs.

These have mainly been centered on mobile device chips but size is size, and I assume it would apply to GPU with INtel having to make these new chips in the near future. Although I guess Intel could possibly kill off the discrete GPU (as it has done with many components after components in its history) by refusing to make them, and in return of its decision, perhaps have more powerful graphic module on its chips at 14-10nm than a discrete GPU at 20nm.


Intel-speak or any truth to it?

 

[NTMBK]

Anyone know about reports that these foundries are having a difficult time at these sizes ( below 20) with extremely low yield potential making it cost prohibited to commercial make these chip sizes for the foreseeable future. I have been reading, in investment reports, the last few years that yield rates be so low on these sizes at most foundries for hire that it will result in delay after delay.

This is the reason Intel has been so aggressive in foundry buildout the last few years (doubling or tripling its entire foundry capacity - it was preparing for something big.) The jest is that for commercial successful yields over the next few years, most all business in these chips sizes will have to go through Intel with Intel's only question being whether if it will become a foundry for hire allowing ARM to survive by building chips under that architecture at those sizes or if it insists on Intel design only, therefore basically killing ARM as far as new chip wins/designs.

These have mainly been centered on mobile device chips but size is size, and I assume it would apply to GPU. Although I guess Intel could kill off the discrete GPU by refusing to make them if only it can make these sizes, and perhaps have more powerful graphic module on its chips at 14-10nm than a discrete GPU at 20nm.


Intel-speak or any truth to it?

Given how massively delayed Intel's 14nm parts have been, and the extremely slow and incremental roll out of Broadwell parts, I suspect that Intel isn't having a great time on 14nm either. (Though clearly still doing significantly better than the foundries!)

 

[RussianSensation]

Wait till GM200 and R9 390X are launched. What would you be saying if AMD recovers the lost market share and gains some more at Nvidia's expense in 2015.

I can't predict what will happen to market share after R9 300 launches, but I sure want to see some progress made with GM200 and 390X because recent game standing of 970/980 verify the disappointment I had 3-4 days after GM204 launch reviews after I took a step back and realized just how underwhelming GM204 was. I was extremely close with ditching my 7970s for Gigabyte 970s G1 but I am glad I didn't fall for the hype.

At high-rez gaming, 980 is only 46-47% faster than 7970Ghz but 3 years later we should expect 70-100% faster. Considering 980 is $550-600, it's mind-boggling how successful that card is given how little it brought over 290X/780Ti and in the context of time against a 3-year-old 7970Ghz, it simply awful, the worst next gen increase in 3 years EVER.

We shouldn't have a situation where 980 SLI is only 30% faster than 1.05Ghz 7970s in modern games since 7970 came out nearly 3 years ago!

I expect nothing less than 70%+ faster than 7970Ghz from GM200/210 and R9 390X.

Why do you think AMD would risk falling behind badly against Nvidia ? You always seem to have a bad prediction waiting for AMD. :biggrin:

1. NV users don't wait for AMD, and AMD gamers who have reached their upgrading point don't either. That's why it's more critical for AMD to be on time than it is for NV. Fermi and Kepler prove that this is true more than anything. NV can be 3-9 months later with its top-to-bottom line-up roll-out and it's OK. When AMD is late by 3-6+ months, this has devastating effects on its market share (2900XT, R9 290/290X series, and now R9 300 series).

2. TSMC has a better track record in execution than GloFo. If AMD switches for GPU manufacturing to GloFo, that to me screams is more risk of delays for AMD's GPUs, which are already behind NV's Maxwell execution. If NV continues to execute at the same pace and so does AMD, AMD will be 5-6 months behind every new generation from NV. This will be devastating for AMD's market share strategy. For AMD to be behind, they can't do a 290/290X repeat. If they are late, their cards should be faster as the market has already shown that when AMD is late, price/performance doesn't work for them. The easiest way to lose the battle is to not show up. AMD seems to be 'great' at this.

 

[raghu78]

I can't predict what will happen to market share after R9 300 launches, but I sure want to see some progress made with GM200 and 390X because recent game standing of 970/980 verify the disappointment I had 3-4 days after GM204 launch reviews after I took a step back and realized just how underwhelming GM204 was. I was extremely close with ditching my 7970s for Gigabyte 970s G1 but I am glad I didn't fall for the hype.

Good decision. :thumbsup: I am sure R9 390X will not disappoint. The last time AMD got first to a new memory standard was HD 4870 and Nvidia did not catch up for 18+ months. The price wars were fierce during the GTX 280/GTX260 and GTX 285/GTX 275 days and consumers benefited. Nvidia's marketshare kept eroding all through that period and continued with HD 5870/HD 5850. AMD hit their all time high and Nvidia hit an all time bottom in Q3 2010.

http://www.xbitlabs.com/news/graphi..._on_Discrete_GPU_Market_Mercury_Research.html

From there onwards Nvidia has done very well and AMD has done reasonably well in desktop but very poorly in notebook. Nvidia commands 65% of the overall GPU market nowadays.

The future looks much better for AMD. The time to market and competitive advantage with HBM will reflect in 2015 market share. Notebooks also benefit hugely as till now MXM could only support GPUs with 256 bit memory. But with 2.5D stacking the notebook GPUs can have massive 4096 bit memory bus with > 400 GB/s and have more than 2.5x the bandwidth of current flagship notebook GPUs

http://en.wikipedia.org/wiki/Mobile_PCI_Express_Module
http://www.mxm-sig.org/file.cfm?doc=36A88883-DFFA-E3E6-16C301BCF5633572

2016 brings FINFET APUs from AMD which will be powered by a new x86-64 core. a quad core with 1024 - 1280 GCN 2.0 cores and HBM is enough to eat into the low end GPU market which brings the majority of volume. With the bandwidth problem addressed AMD APUs will perform just like their discrete GPU equivalents.

AMD's time to market advantage with HBM will manifest in desktop and notebook GPUs first and then with 2016 FINFET APUs. AMD is going to eat a lot of low end notebook GPU marketshare which is where the bulk of volume is.

At high-rez gaming, 980 is only 46-47% faster than 7970Ghz but 3 years later we should expect 70-100% faster. Considering 980 is $550-600, it's mind-boggling how successful that card is given how little it brought over 290X/780Ti and in the context of time against a 3-year-old 7970Ghz, it simply awful, the worst next gen increase in 3 years EVER.

We shouldn't have a situation where 980 SLI is only 30% faster than 1.05Ghz 7970s in modern games since 7970 came out nearly 3 years ago!

I expect nothing less than 70%+ faster than 7970Ghz from GM200/210 and R9 390X.

1. NV users don't wait for AMD, and AMD gamers who have reached their upgrading point don't either. That's why it's more critical for AMD to be on time than it is for NV. Fermi and Kepler prove that this is true more than anything. NV can be 3-9 months later with its top-to-bottom line-up roll-out and it's OK. When AMD is late by 3-6+ months, this has devastating effects on its market share (2900XT, R9 290/290X series, and now R9 300 series).

2. TSMC has a better track record in execution than GloFo. If AMD switches for GPU manufacturing to GloFo, that to me screams is more risk of delays for AMD's GPUs, which are already behind NV's Maxwell execution. If NV continues to execute at the same pace and so does AMD, AMD will be 5-6 months behind every new generation from NV. This will be devastating for AMD's market share strategy. For AMD to be behind, they can't do a 290/290X repeat. If they are late, their cards should be faster as the market has already shown that when AMD is late, price/performance doesn't work for them. The easiest way to lose the battle is to not show up. AMD seems to be 'great' at this.

I agree TSMC has a better track record in execution than Samsung and GF. I also believe TSMC 16FF+ will be the best foundry process in terms of performance, power,area and yields when compared against Samsung 14LPE/14LPP. But I don't expect AMD to be late to FINFET. If anything AMD has always been the first to a node. This has been the case from 55nm down to 28nm.

I expect AMD to split their FINFET products across both GF and TSMC to avoid capacity/supply problems. So in places where high performance is required I expect AMD to go with TSMC 16FF+ - like server/desktop enthusiast CPUs and GPUs. In places where cost is more important than raw performance they might go with GF eg:- APUs, game console and semi-custom chips . Samsung / GF are looking to undercut TSMC on wafer prices. But as always that makes sense only if yields are reasonably competitive. :thumbsup:

 

[krumme]

yeah. I am quite sure that AMD has gone with GF 28SHP and 2.5D stacking on silicon interposer with HBM.

http://seekingalpha.com/article/272...redit-suisse-technology-conference-transcript

Devinder Kumar - SVP and CFO

"I can tell you that with the changes that have occurred on GlobalFoundries with the management team and the focus that Abu Dhabi has and the investment that they have as a partnership with GlobalFoundries, the relationship between AMD and GlobalFoundries is the best in the history of the relationship. The folks that we're working on in fact we just had a meeting in Abu Dhabi just a couple of weeks ago. Really good discussions, very business oriented.

And I think the execution of GlobalFoundries has improved significantly and that helps us from an overall standpoint. In 2014 for the first time, some folks may not know then. For the first time in the history of the relationship we went beyond PC product and actually we are making graphics, PC, and semi-custom products at GlobalFoundries in 2014 and that continue into 2015. When you diversify the product that you make at a foundry like GlobalFoundries, it benefits them from a mix standpoint and benefits us from a mix standpoint. And like I said, the execution is continuing to get better and we are very pleased - very, very pleased with that relationship."

AMD is already manufacturing Kaveri, Beema, Mullins, game console chips and graphics at GF 28SHP. For 2015 I expect AMD to go top to bottom with GF 28SHP - GPUs, APUs (desktop,notebook,embedded), semi-custom game console chips, Seattle server SoC and any other newer server chips if planned.

Given the fact that GF has licensed Samsung 14nm FINFET and will be ramping production in H1 2015 (Samsung expects to start 14nm FINFET volume production by end of 2014.) I wouldn't be surprised if AMD is one of the first adopters of Samsung/GF 14nm FINFET. We can expect 14nm FINFET GPUs from AMD by late 2015 or early 2016.

The Hector Ruiz deal put both gf and especially amd in a hopeless unsecure, clouded situation. Amd topmanagement can nothing but bend and talk positively.

Now Kumar actively say something about the relationship. And read between the lines its obvious it have been extremely difficult. I only interprete it positively. Otherwise he would just shut up.

It cant possible be worse but its a professional failure that the execution of gf should alter the quality of the relationship. Perhaps way to many backdoors or unclear situation in the original arangement. Rushed work and not in anyones interest.

Dont give topmangememt to much bonus for selling off part of the company.

 

[NostaSeronx]

AMD will absolutely go bankrupt if they went FinFETs. Sub-300 million R&D per quarter with a completely new structure to get use to. Sub-par to more than sub-par devices from AMD till Chapter 7.

 

[krumme]

AMD will absolutely go bankrupt if they went FinFETs. Sub-300 million R&D per quarter with a completely new structure to get use to. Sub-par to more than sub-par devices from AMD till Chapter 7.

Well for two and a half years crolles have been "ready" with 28nm fdsoi. Yet nothing happens. And If/when it happens end 2015 h1 its like 3 years late. Then what is it worth? At that time "cheaper" finfet process tech is near anyway.