[WCCF] AMD Contracts TSMC To Produce Zen At 16nm Amidst Concerns Over 14nm

[frozentundra123456]

The Intel process roadmap you showed was talking about production start, not product availability. Typical fab cycle time is 1-1.5 quarters, so production start in Q2 2014 means products on shelves in late Q3 2014.

Again, a two quarter delay is an "oopsie," but the outright cancellation of a node that was expected to arrive in what seemed like an overly aggressive timeline is indicative of an attempt to deceive, IMO.

Maybe, but the "oopsie" came at the worst possible time, while they were desperately trying to get a foothold in mobile. They needed a compelling product on time, and instead got a mediocre, or at least less than compelling, product that was late and still doesn't seem to have the yield issues completely sorted out. Believe me, I would love for intel to kick some ARM butt and get a strong presence in the phone market and in the tablet market without contra revenue, but I am afraid they have missed their opportunity, especially since 10nm is going to be delayed as well.

But like I said, Zen's hopes are tied to x86 as well, so I dont think AMD fans should be taking satisfaction in intel's struggles, because if ARM makes inroads into the server market, AMD will be fighting two or more very strong competitors instead of just intel.

 

[JDG1980]

Is there any semiconductor manufacturer that hasn't gotten delays on 14/16 nm?

All of them were delayed, but some worse than others.

The first Core M laptop, and thus the first Intel 14nm product on the market, was released about one year ago, October 2014.

Samsung came to market with the Galaxy S6 in April 2015. This contains the Exynos 7420 SoC, making it the first shipping product on the Samsung 14nm FinFET process (presumably LPE).

The iPhone 6S and 6S Plus were just released this month (September 2015). There are two different versions of this phone's A9 SoC, one manufactured at Samsung and one at TSMC. This is therefore TSMC's first shipping 16nm FinFET product.

Global Foundries? Nothing so far except worthless PowerPoint slides. And the fact that they're desperate enough to try to reuse old equipment should not fill anyone with confidence about their ability to get it up and running.

What exactly does the WSA require? Does it put any obligations on GloFo to deliver in a timely manner? GloFo has already cost AMD quite a bit of money due to the inability to get 20nm working. Worthless 20nm tape-outs used up the R&D funds that should have gone to a 28nm GPU refresh. AMD is betting the future of the company on Zen, and it would be monumentally foolish to let this rest on GloFo's shoulders. (Of course, the fact that something is monumentally foolish has not, historically, been an impediment to AMD doing it.) AMD really ought to go with TSMC for Zen; whatever delays they've had in the past finally seem to be resolved, and the 16FF+ process is actually designed for high-power applications, while both of Samsung's processes are really smartphone-focused. In contrast, GloFo may never deliver at all, and if they do, it will inevitably be late and the performance will be disappointing.

 

[Arachnotronic]

Maybe, but the "oopsie" came at the worst possible time, while they were desperately trying to get a foothold in mobile. They needed a compelling product on time, and instead got a mediocre, or at least less than compelling, product that was late and still doesn't seem to have the yield issues completely sorted out.

I agree, Cherry Trail is a disappointment.

 

[RussianSensation]

AMD really ought to go with TSMC for Zen; whatever delays they've had in the past finally seem to be resolved, and the 16FF+ process is actually designed for high-power applications, while both of Samsung's processes are really smartphone-focused. In contrast, GloFo may never deliver at all, and if they do, it will inevitably be late and the performance will be disappointing.

GloFo has continuously has failed to meet its timelines. If GLoFo's Samsung derived 14nm FinFET is similar to Samsung's, then in terms of die size there is some benefit though:

Since AMD likely cannot afford dual-sourcing, they better start thinking really hard about a possible impact of missing 2-3 quarters schedule and sticking with GloFo. I hope their lawyers were smart enough to have clauses in there that if GloFo cannot deliver tape-outs on time, then AMD is free to opt out of the WSA. It could be far more difficult though since Mubadala is a joint shareholder in both AMD and GloFo so they could have significant influence on AMD's management here.

I haven't had time to do more up-to-date research but last time I checked:

"Mubadala Development Co. extended its stake in AMD to about 19 percent including warrants and took an extra board seat last month"

and GloFo will bend over for Mubadala.

"Abu Dhabi's Advanced Technology Investment Co (ATIC) plans to invest up to $10 billion over the next two years in GlobalFoundries' upstate New York semiconductor factory, its chief executive said on Friday.

ATIC owns unlisted GlobalFoundries, having completed a buyout of joint venture partner Advanced Micro Devices Inc in March 2012. ATIC is controlled by Abu Dhabi state investment fund Mubadala." ~ Reuters

--> Actually it's MUCH more complicated than I thought:

"GLOBALFOUNDRIES, which is [now] 100% owned by Mubadala, has the simple but ambitious goal of becoming the world’s first truly global semiconductor manufacturing company." ~ Source

"AMD, in which Mubadala holds a significant equity stake, designs and integrates technology that powers millions of intelligent devices. It is the world’s second largest supplier of microprocessors and the third largest supplier of graphics processing units.

AMD’s server computing products are focused on driving industry-leading cloud computing and virtualization environments. Its graphics technologies are found in a variety of solutions ranging from game consoles to personal computers and supercomputers." ~ Source

It's going to be interesting to see how AMD's management can solve this huge dilemma. That's also no wonder that Intel continues to tout that its cutting-edge manufacturing is a key sustainable & competitive advantage at every node. Without its own cutting edge fabs for CPUs, APUs and GPUs, AMD is always at the mercy of 3rd parties (Samsung, TSMC, GloFo, etc.). That's actually one of the key reasons why even if AMD caught up to Intel in IPC, unless they can manufacture their CPUs on a node that's as good as Intel's, they won't beat Intel in perf/watt either without actually making a CPU architecture that's superior in perf/watt by overcoming the massive manufacturing advantage Intel has!

 

[raghu78]

GloFo has continuously has failed to meet its timelines. If GLoFo's Samsung derived 14nm FinFET is similar to Samsung's, then in terms of die size there is some benefit though:

GF uses a copy exact of Samsung 14nm and if Apple has chosen Samsung to power the iPhone 6S (APL0898 was in 6S) then I doubt there are yield problems. There is still a bit of confusion on whether APL1022 manufactured by TSMC is used only in 6S Plus. If that were true it would clearly dismiss the Samsung / GF yield problem rumours as 6S unit sales volume is two times that of 6S Plus.

Since AMD likely cannot afford dual-sourcing, they better start thinking really hard about a possible impact of missing 2-3 quarters schedule and sticking with GloFo. I hope their lawyers were smart enough to have clauses in there that if GloFo cannot deliver tape-outs on time, then AMD is free to opt out of the WSA. It could be far more difficult though since Mubadala is a joint shareholder in both AMD and GloFo so they could have significant influence on AMD's management here.

AMD must have clauses on yields in their WSA to hold GF accountable. If GF does not deliver agreed upon yields AMD should be free to move Zen to TSMC. Thats what I consider a reasonable agreement. GF has been harming AMD for too long and AMD should not allow this to continue.

 

[frozentundra123456]

What has "reasonable" ever had to do with the toxic relationship between GF and AMD?

And even if you are correct that there are no yield problems with the iphone, I dont think that because they have a good yield of those that it rules out that there could still be yield problems with 8 core high performance chips like Zen. It is a completely different animal than phone chips.

 

[Fjodor2001]

Even WccfTech seems to have retracted from the rumour in the OP now:

http://wccftech.com/globalfoundries-14nm-finfet-amd-zen-2016/

Globalfoundries 14LPP Is On Time For AMD’s Zen – Volume Ramp To Start In Early 2016

[2015-09-28]
Globalfoundries communicated the progress of its 14nm plans this past Saturday, asserting that its 14LPE FinFET process is on track for volume production this year, while its high performance 14LPP is set for qualification in the second half of the year with volume ramp set to take off early next year.

A rumor emerged last week claiming that AMD had relocated the production of its upcoming Zen based chips from Globalfoundries to TSMC. Which allegedly occurred as a result of concerns from the Sunny Vale California based chip maker over sluggish progress of Gloablfoundries’ 14nm FinFET process plans. A setback that was allegedly fueled by a challenging economic climate in the Emirate of Abu Dhabi, which owns Globalfoundries. We should note that none of these claims were verified or validated, and as such this is still very much a rumor.

However, speaking to WCCFTech Jason Gorss, Senior Manager, Corporate and Technology Communications at Globalfoundries affirmed this past Saturday that progress for the company’s 14nm FinFET manufacturing is not only on track but is actually ahead of schedule and 14nm has so far managed to exceed plans for yield and defect density, stating :

Our 14nm FinFET ramp is exceeding plan with best-in-class yield and defect density. The early-access version of the technology (14LPE) was qualified in January and is well on its way to volume production, meeting yield targets on lead customer products. The performance-enhanced version of the technology (14LPP) is set for qualification in the second half of 2015, with the volume ramp beginning in early 2016. Prototyping on test vehicles has demonstrated excellent logic and SRAM yields and performance at near 100% of target.

[...]
I should clearly point out that AMD has refused on multiple occasions to specify where exactly it plans to produce its Zen based products launching next year. It maintains that these parts will be produced on “FinFET” but does not provide much detail beyond that. Globalfoundries also doesn’t talk about specific client products or relationships. And while AMD still does not want to officially commit to any one particular foundry for the production of its hotly anticipated Zen based CPUs, Globalfoundries remains a likely candidate.

I still think the statements from GF are a bit vague. They are mostly focusing on saying that their 14LPE and 14LPP process techs are progressing fine. But I'd like to see a clear denial of that AMD has moved their Zen production to TSMC. However the last two sentences in bold above perhaps explains why they cannot say more than this...

 

[JDG1980]

GloFo has continuously has failed to meet its timelines. If GLoFo's Samsung derived 14nm FinFET is similar to Samsung's, then in terms of die size there is some benefit though

That's a very small difference, and not worth it if it means even a marginal reduction in maximum clock rate. Assuming IPC falls somewhere between Sandy Bridge and Haswell levels, Zen needs to hit at least 3.5 GHz to be competitive on the desktop, and 4 GHz or more would be ideal. (Server CPUs can get away with lower clock speeds.) That's going to be very tough, and may be impossible on a process designed primarily for smartphones. At least TSMC's 16FF+ is designed with high-power applications in mind.

Since AMD likely cannot afford dual-sourcing, they better start thinking really hard about a possible impact of missing 2-3 quarters schedule and sticking with GloFo. I hope their lawyers were smart enough to have clauses in there that if GloFo cannot deliver tape-outs on time, then AMD is free to opt out of the WSA. It could be far more difficult though since Mubadala is a joint shareholder in both AMD and GloFo so they could have significant influence on AMD's management here.

I think this is the core of the problem. Not only is the WSA a turd and a millstone around AMD's neck, but the lack of an arm's-length relationship between the two companies means that AMD management is unwilling to play hardball with GloFo.

As things currently stand, AMD has all the disadvantages of being fabless (lack of vertical integration, lack of control over when new processes are developed and what they're optimized for), and none of the advantages. This could be fatal.

 

[JDG1980]

GF uses a copy exact of Samsung 14nm

The Expreview/WCCFTech article indicates that this is not the case; GloFo is trying to save money by repurposing old 28nm equipment instead of copying Samsung's process for 14nm. I don't see this working out well. It is likely to result in substantial delays and yield problems at the very least, and may never work at all. GloFo's press releases have no credibility at this point; I won't believe they have 14nm FinFET up and running until I see an actual shipping product.

and if Apple has chosen Samsung to power the iPhone 6S (APL0898 was in 6S) then I doubt there are yield problems. There is still a bit of confusion on whether APL1022 manufactured by TSMC is used only in 6S Plus. If that were true it would clearly dismiss the Samsung / GF yield problem rumours as 6S unit sales volume is two times that of 6S Plus.

There is no basis for this assertion. Chipworks describes finding "two different application processors in two otherwise identical phones". If one was a 6S and one was a 6S Plus then they would not be "otherwise identical".

AMD must have clauses on yields in their WSA to hold GF accountable. If GF does not deliver agreed upon yields AMD should be free to move Zen to TSMC. Thats what I consider a reasonable agreement. GF has been harming AMD for too long and AMD should not allow this to continue.

No one seems to know what the WSA actually says. I've found several revisions of it floating around on the Internet, but all are heavily redacted. It's not clear if there are any performance obligations for new nodes on GloFo's part, or if AMD is obligated to shovel money down GloFo's maw no matter how badly they underperform.

 

[mrmt]

GF uses a copy exact of Samsung 14nm and if Apple has chosen Samsung to power the iPhone 6S (APL0898 was in 6S) then I doubt there are yield problems. There is still a bit of confusion on whether APL1022 manufactured by TSMC is used only in 6S Plus. If that were true it would clearly dismiss the Samsung / GF yield problem rumours as 6S unit sales volume is two times that of 6S Plus.

You can't dismiss yields problems just because of the A9 volumes. Samsung could be with bad yields but they could be biting the bullet and still delivering by adding more wafers to the mix. Something akin to what TSMC did with NVidia on the beginning of the 28nm life cycle, when they paid for good chip and not per wafer.

 

[AtenRa]

That's going to be very tough, and may be impossible on a process designed primarily for smartphones. At least TSMC's 16FF+ is designed with high-power applications in mind.

ZEN will use the GloFo 14nm LPP which is for high performance applications.

 

[Atreidin]

You can't dismiss yields problems just because of the A9 volumes. Samsung could be with bad yields but they could be biting the bullet and still delivering by adding more wafers to the mix. Something akin to what TSMC did with NVidia on the beginning of the 28nm life cycle, when they paid for good chip and not per wafer.

I would actually be surprised if Apple didn't negotiate a similar contract, where they only pay for good chips. It sounds like the kind of thing they would do. I would be interested in finding out.

 

[dark zero]

It will end into a MASSIVE dissaster... The Core is great... GloFo is the problem now.

And Nvidia will suffer too not because their.good chips... They will suffer due TSMC hiatuses... And... If their Apple A9 at 16 nm is inferior than Samsung at 14 nm by a margin of 10% or more... We might get an idea on how good Nvidia will perform.