Samsung and Global Foundries will produce AMD's Next Gen Greenland GPU and Zen CPU

[VR Enthusiast]

Don't we already have an AMD slide showing that 14nm is cheaper per transistor?

Does 16FF even exist for TSMC now? I can't find any evidence of it in their materials, only 16FF+ and 16FFC (compact) coming next year.

Chipworks i6s teardown hints that the A9 could be 14 LPE with Samsung and they are still not sure whether it is FF or FF+ with TSMC.

 

[ShintaiDK]

16FF+ is in mass production for quite a while.

A9, A9X, Xilinx FPGA etc is made on it.

 

[Arachnotronic]

16FF+ is in mass production for quite a while.

A9, A9X, Xilinx FPGA etc is made on it.

Kirin 950, too.

 

[VR Enthusiast]

Ok well this could be really bad news for Nvidia because it very much looks like the A9 is LPE and FF+.

If TSMC's 16FF+ is only just capable of beating Samsung's 14 LPE with around 9% area disadvantage then it should be much better for AMD with LPP being what 15% better than LPE?

The weak NMOS of the LPE is the reason why the Samsung A9 drains more battery at constant maximum load, but this should be solved with LPP. LPP could need more area though, equalizing that with TSMC.

Nvidia should have a TTM advantage but that will depend on Pascal being ready, not the FF+ process. Pascal only taped out in June, or GP100 did, so it is likely that both AMD and Nvidia will be releasing their new GPU in mid 2016 earliest, yet AMD with the superior process. That is what makes most sense based on the infos we have.

Note that the TSMC A9 could be capable of better performance too but is being very conservatively binned due to process maturity and Apple needing to keep both chips within a close performance envelope. The results we've seen do appear to suggest that the TSMC A9 is built on a later, more high-performance process vs the earlier (higher yield) Samsung process that isn't meant for high performance.

 

[ShintaiDK]

A9 is 14LPP and 16FF+.

Also A10 will be TSMC only using their 16nm node.

Area isn't going to give AMD any benefits, we know that from previous experience. And 16FF+ is superior to 14LPP in electrics. If anything, it looks like another disadvantage to AMD.

 

[VR Enthusiast]

A9 is 14LPP and 16FF+.

What makes you believe that it is 14 LPP instead of 14 LPE?

Also A10 will be TSMC only.

This is not relevant to the discussion.

Area isn't going to give AMD any benefits, we know that from previous experience.

What previous experience? Maxwell? In every other case I can remember in the last 7 years AMD has been faster or really close with much smaller dies.

And 16FF+ is superior to 14LPP in electrics

TSMC's A9 appears to be slightly better than Samsung's A9 under constant high-load conditions. However, it is much more likely to be an LPE chip than an LPP chip based on what we know and what most in the industry say.

 

[ShintaiDK]

What makes you believe that it is 14 LPP instead of 14 LPE?

Several sources and Samsungs own statement about 14LPP mass production.

This is not relevant to the discussion.

Why not, because it tells a grim story you dont want to hear?

 

[VR Enthusiast]

Several sources and Samsungs own statement about 14LPP mass production.

Can you provide proof of any source stating that Samsung was mass producing A9's on LPP?

Why not, because it tells a grim story you dont want to hear?

No it's because the A10 production has zero relevance to the A9 production. There are lots of reasons why TSMC are the better choice for Apple that have nothing to do with performance.

 

[ShintaiDK]

You are asking for something nobody can deliver because there isn't an official statement from Samsung or Apple. Just like you cant deliver the same with your 14LPE argument.

A9X is TSMC exclusive too. But its also a higher performance part.

 

[Arachnotronic]

You are asking for something nobody can deliver because there isn't an official statement from Samsung or Apple. Just like you cant deliver the same with your 14LPE argument.

It will be a cold day in hell when Apple or its suppliers talk about these sorts of details publicly.

Bottom line is that TSMC's transistor performance with 16FF+ is generally better than Samsung's 14LPP and yields are said to be much, much higher on the TSMC A9 than on the Samsung one.

A lot of people here (particularly AMD supporters) seem to be cheerleading for Samsung because this is the process that AMD will be using.

Anyway just sit back, relax, and watch as the vast majority of fabless customers choose TSMC 16FF+. It will be hard to spin it once the reality is obvious to everybody

 

[VR Enthusiast]

You are asking for something nobody can deliver because there isn't an official statement from Samsung or Apple. Just like you cant deliver the same with your 14LPE argument.

A9X is TSMC exclusive too. But its also a higher performance part.

I can deliver multiple, trustworthy industry sources stating or suggesting it is likely to be 14 LPE. The only "sources" contrary to this view are you and Arachnotronic...

Bottom line is that TSMC's transistor performance with 16FF+ is generally better than Samsung's 14LPP and yields are said to be much, much higher on the TSMC A9 than on the Samsung one.

Yields are much higher too now? Yet 2/3rds of A9's are Samsung built and TSMC has vastly more capacity...how does that work?

Very much doesn't add up with your theories. The common sense argument is that Apple was forced to go with Samsung because TSMC was very late. It does appear that TSMC has made an ok "+" process - barely better than Samsung's LPE - but that is enough for Apple who would rather not work with Samsung so long as TSMC can deliver anythign workable.

 

[USER8000]

Has anyone stopped to ponder,that BOTH Nvidia and AMD uarchs are very different and that different process nodes might actually work better for each of them. The AMD GCN cards run at a much lower sustained clockspeed than Maxwell and even to some degree Kepler did,so they will probably be going for lower clockspeeds and more shaders if the process is denser but is less efficient. Nvidia will probably go for higher sustained clockspeeds,with less shaders so a more efficient node will probably fit their way of doing things.

 

[Hans de Vries]

It will be a cold day in hell when Apple or its suppliers talk about these sorts of details publicly.

Bottom line is that TSMC's transistor performance with 16FF+ is generally better than Samsung's 14LPP and yields are said to be much, much higher on the TSMC A9 than on the Samsung one.

A lot of people here (particularly AMD supporters) seem to be cheerleading for Samsung because this is the process that AMD will be using.

Anyway just sit back, relax, and watch as the vast majority of fabless customers choose TSMC 16FF+. It will be hard to spin it once the reality is obvious to everybody

Apple's A10 can not be exclusively produced by TSMC, which slashed its
capex by 30% due to lost opportunities in the mobile sector... To late now.
10nm is the next opportunity, the next race.

Apple's A9 is produced on 14LPE by both Samsung and Globalfoundries.
14LPE has reached "world class yields" according to very happy managers
like Garry Patton (Globalfoundries, ex-IBM) a trusted and respected veteran.

For Q4 Apples uses a gargantuan amount of silicon (more than Intel
produces). Should we believe that Samsung can ramp up 14LPP on day
one from 0 to >70% of Intel's production capacity and do so even with
very crappy yields?

Simple industry facts and logic versus "FUD re-posted every 10 minutes
during 18 hours per day by the same anonymous poster day in and day out"

And now for the more interesting rumors:

The "Samsung" version of the A10 will be on 14LPH. One of three new
processes 14LPC, 14LPA and 14LPH (Compact, Advanced, High perf.)
presumably produced in tandem with Globalfoundries like they do now.

 

[Arachnotronic]

Apple's A10 can not be exclusively produced by TSMC

Serious question: how much would you like to bet on that?

 

[Abwx]

Has anyone stopped to ponder,that BOTH Nvidia and AMD uarchs are very different and that different process nodes might actually work better for each of them.

No, a transistor is a transistor, a process is a methodology to manufacture transistors.

The AMD GCN cards run at a much lower sustained clockspeed than Maxwell and even to some degree Kepler did,so they will probably be going for lower clockspeeds and more shaders if the process is denser but is less efficient. Nvidia will probably go for higher sustained clockspeeds,with less shaders so a more efficient node will probably fit their way of doing things.

Nvidia paid to have a higher grade process, like Altera before they were bought by Intel , not counting that TSMC are playing dirty with AMD on the behalf of Nvidia..

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

Seriously, when was 28nm released, and theres s still yields issues...?

They are sabotaged willfully, it doesnt matter for TSMC that AMD sell more as what is not sold by AMD will be sold by Nvidia...

 

[Arachnotronic]

Nevermind.

 

[Hans de Vries]

Serious question: how much would you like to bet on that?

The Capex reduction excludes it, and Warren Lau has a good track record:

source:

'TSMC will receive one-third of Apple's A9 allocation and half of its A10 allocation,
' said Warren Lau, an analyst with Maybank Kim Eng, in an August 24 report.
'An assumption of 50% is reasonable given Apple's desire to maximize its bargaining
power with suppliers, 'Lau said.

 

[Arachnotronic]

The Capex reduction excludes it, and Warren Lau has a good track record:

Doesn't answer my question

Anyway, Warren Lau in this (old) note was just guessing. There have been multiple independent (and public, so you don't have to take my word for it) sources that have said that A10 will be 100% TSMC including:

1. KGI Securities' Ming-Chi Kuo (http://www.macrumors.com/2015/11/03/kuo-4-inch-iphone-2016-iphone-7-ram/)

2. HSBC analysts (http://blogs.barrons.com/asiastocks...egain-apple-in-a10-info-hsbc-upgrades-to-buy/)

3. China Times (http://www.chinatimes.com/newspapers/20150914000039-260202)

4. Last, but IMO the most reliable source, frankly, is our own Idontcare (http://forums.anandtech.com/showpost.php?p=37716711&postcount=34). This poster correctly pointed out on these forums that TSMC had won the entirety of A8 when the rest of the world was saying it would be split, and said that A9 would be split between TSMC/Samsung even though many were saying TSMC couldn't possibly be involved and/or splitting the orders on a single design made no sense. Also note in the linked post, Idontcare said that A9X was 100% TSMC which was later confirmed when iPad Pro hit the market.

 

[el etro]

Nvidia paid to have a higher grade process, like Altera before they were bought by Intel , not counting that TSMC are playing dirty with AMD on the behalf of Nvidia..

Seriously, when was 28nm released, and theres s still yields issues...?

They are sabotaged willfully, it doesnt matter for TSMC that AMD sell more as what is not sold by AMD will be sold by Nvidia...

Where do you got this Abwx? Because it makes much sense on AMD dumping TSMC.

 

[JDG1980]

The difference between the Samsung and TSMC A9 chips is negligible, based on a series of tests run by Tom's Hardware. Most other sites came to a similar conclusion; any difference between the two chips is a few percent at most. Apple has publicly stated that variance will be only 2 to 3 percent. I think it is safe to say that whether Pascal or Arctic Islands comes out ahead in performance and perf/watt is going to depend on the architectures, and not on which FinFET process the companies choose to use.

It's a mistake to think that because Apple prefers TSMC, that proves the TSMC process is better. The truth is that Apple would prefer not to work with Samsung, who is their biggest competitor in the smartphone business. As things currently stand, they have no choice but to use Samsung for some of their A9 business because TSMC's yields on the new process likely weren't high enough to meet Apple's production targets alone. Once the process has all its kinks worked out (and this will happen, just like with every other process node before), then Apple is naturally going to go with the company they're not embroiled in commercial and legal battles with. That in no way indicates that TSMC > Samsung for FinFETs in some absolute technical sense.

 

[Hans de Vries]

4. Last, but IMO the most reliable source, frankly, is our own Idontcare (http://forums.anandtech.com/showpost.php?p=37716711&postcount=34). This poster correctly pointed out on these forums that TSMC had won the entirety of A8 when the rest of the world was saying it would be split, and said that A9 would be split between TSMC/Samsung even though many were saying TSMC couldn't possibly be involved and/or splitting the orders on a single design made no sense. Also note in the linked post, Idontcare said that A9X was 100% TSMC which was later confirmed when iPad Pro hit the market.

But the good man also claimed (well, of course his sources claimed)
that TSMC was doing almost all of the A9 because Samsung and
Globalfoundries screwed up completely with their yields. That the A9X
would be 100% was already know. Also what he said was before the
large 2015 capex reduction of TSMC in October.

Having said that, with 14LPP never ever on the road map for production
before 2016 for Globalfoundries, it could only have had delays with an
14LPE A9 ramp.

 

[Arachnotronic]

But the good man also claimed (well, of course his sources claimed)
that TSMC was doing almost all of the A9 because Samsung and
Globalfoundries screwed up completely with their yields. That the A9X
would be 100% was already know. Also what he said was before the
large 2015 capex reduction of TSMC in October.

The capex reduction doesn't seem to have much to do with 16FF share shifts.

Here's what TSMC management attributed it to:

Of this change here, about 33% or one-third of the reduction of CapEx is due to operating efficiency gains that led us to spend less money but still has the same output. Another one about 30% of the reduction is due to changes in investment projects including the conversion between 20 nanometer and 16 nanometer. Another one, about 20% of the reduction is due to changes in capacity schedules. Lastly, the remainder about 17% of the total reduction is due to the strengthened of U.S. dollars against euro, Japanese yen and NT dollars. We expect our 2016 capital budget will be higher than this year.

17% of the reduction was due to forex, 33% was due to increased efficiency, so right off the bat about half of the reduction had nothing to do with capacity/market share.

The remaining capex reduction doesn't scream "we lost big time to Samsung"

 

[Hans de Vries]

The remaining capex reduction doesn't scream "we lost big time to Samsung"

The 2015 capex was reduced in two steps from almost 12 billion to
8 billion dollar. That doesn't scream "we won big time from Samsung"
either.

Now, I like TSMC a lot, so lets hope they prosper both and provide
us with the best possible technology in the years to come.

 

[Arachnotronic]

The 2015 capex was reduced in two steps from almost 12 billion to
8 billion dollar. That doesn't scream "we won big time from Samsung"
either.

Well, I'm not so sure. The reasons they gave were around timing of capacity as well as changes in 20nm->16nm migration plans. I wonder if they meant that they are taking 20nm capacity offline and retooling for 16nm more aggressively than planned as demand for 20nm wanes?

Now, I like TSMC a lot, so lets hope they prosper both and provide
us with the best possible technology in the years to come.

We find common ground here. TSMC is probably one of the most under-appreciated semiconductor companies out there, IMO, and I think they will do very well in the years ahead. :thumbsup:

 

[prtskg]

Samsung announced that LPP was already going through volume production in august with several customers. That means it began before that date.

The released Xilinx Ultrascale+ is also based on TSMC 16FF+.
https://www.semiwiki.com/forum/content/5063-xilinx-beats-altera-first-finfet-fpga.html

No, we would not already have seen new GPUs due to cost. NVidia and AMD cant afford to pay the same as others. And unlike previous there is no cost reduction in transistor cost.

Samsung started on A9 much earlier.
http://www.forbes.com/sites/ewanspence/2014/12/13/apple-a9-chip-for-iphone-7-made-in-america/
http://www.technologytell.com/apple/144103/samsung-begins-apple-a9-system-chip-production/
That's why A9 couldn't have been produced on lpp.
Tsmc own date puts ff+ volume production in Q3 2015.
http://www.taipeitimes.com/News/biz/archives/2015/08/11/2003625049

TSMC started production of A9 towards the end of Q2 or beginning of Q3 -
http://www.gforgames.com/gadgets/tsmc-apple-a9-mass-production-june-47581/