Samsung to Fab AMD "Zen" and "Arctic Islands" on its 14 nm FinFET Node

[Ferrum Master]

Back then UMC was still had a seat at the feast (from memory their 80nm was used by ATI/AMD, and their 65nm/55nm was interchangeable with TSMC's - a lot of Nvidia G200/G92's came out of UMC)

It seems we are older than some others. Yes it is true. Same GPU's were made from different suppliers. And obviously one was better.

 

[TRWOV]

Q3 2016 is too far away IMO

 

[midnightoil]

That should be readily apparent I would have thought given that Apple's A9 can be easily ported between TSMC's CLN16FF and Samsung's 14nmLPE. Presumably the same design rules apply for 16nmFF+/FFC and 14nmLPP.

Still have to tape out two different chips, and as tests have proved, they don't share the same electrical properties (the A9 chips). I don't think there's anything easy or cheap about it .....

Previous poster was also wrong. IBM and GF licensed Samsung 14nm FF ... but IBM never implemented it, and GF now own the IBM foundry business anyway. Also, GF 14nm FF LPP isn't copy exact .. it was originally supposed to be but they used different tools to save money ... hence delays and probably yield / volume / cost issues which got AMD to finally make the leap. Furthermore TSMC has absolutely nothing to do with it. They plow their own (lone) furrow.

TSMC 16nm FF+ and Samsung / GF 14nm FF LPP are completely different. FF+ is a high power process. LPP is a low power (but not SLP) process. So AMD / NVIDIA can't swap around without huge cost, redesigns and headaches ... also this predicates a situation where there is finally a major difference in lithography between AMD and NVIDIA - low power process vs high power process; former is likely to have way higher yields and lower costs .. latter will probably clock higher but suffer badly in cost and yields ... and TSMC's FF+ process will have very, very limited / contested capacity at first, unlike Samsung LPP. Samsung's high power (and refined versions of LPE / LPP) come at the very end of '16, though I'd guess the LPE / LPP revisions will take priority over the HP version as AMD have gone with LP designs.

Q3 2016 is too far away IMO

Carrizo desktop and AM4 will probably be launched and available in April or May. Zen likely launches at Computex with availability as soon as they can get it afterwards - some time in Q3. Arctic Islands at E3 with ~immediate availability.

I think the original plan was for Zen desktop (non-APU) to launch with AM4, before AI, so that people could build a new system before the new graphics cards. It'll be the other way around now.

Either way, that isn't long. Especially when neither NVIDIA nor Intel have anything coming up in the mean time.

 

[Steevo]

Why?

I like to think they have some insider knowledge of Samsung.... but then my daydream stops and I realize its a traditional shitpost, none of that new fangled underhanded troll shitpost, just a good old fashioned shitpost, reminiscent of the days where people knew nothing other than what they read somewhere once while getting the old ocular penetration from their favorite phallus.


Also, the news is interesting and should have good effects on the performance/watt metric.

 

[TheLostSwede]

Previous poster was also wrong. IBM and GF licensed Samsung 14nm FF ... but IBM never implemented it, and GF now own the IBM foundry business anyway. Also, GF 14nm FF LPP isn't copy exact .. it was originally supposed to be but they used different tools to save money ... hence delays and probably yield / volume / cost issues which got AMD to finally make the leap. Furthermore TSMC has absolutely nothing to do with it. They plow their own (lone) furrow.

TSMC 16nm FF+ and Samsung / GF 14nm FF LPP are completely different. FF+ is a high power process. LPP is a low power (but not SLP) process. So AMD / NVIDIA can't swap around without huge cost, redesigns and headaches ... also this predicates a situation where there is finally a major difference in lithography between AMD and NVIDIA - low power process vs high power process; former is likely to have way higher yields and lower costs .. latter will probably clock higher but suffer badly in cost and yields ... and TSMC's FF+ process will have very, very limited / contested capacity at first, unlike Samsung LPP. Samsung's high power (and refined versions of LPE / LPP) come at the very end of '16, though I'd guess the LPE / LPP revisions will take priority over the HP version as AMD have gone with LP designs.

Uhm, dude, I was not wrong, this is from about 4-5 years ago and it was announced at a GloFo event. It was before 28nm was mainstream. No-one licensed nothing, it was a shared platform for IC development that could be used to produce the chips with any of the four companies and it was a big deal, as this had never been done before. Looks like Intel was involved as well. It was/is called the Common Platform Alliance.

 

[Fluffmeister]

Either way, that isn't long. Especially when neither NVIDIA nor Intel have anything coming up in the mean time.

It really isn't that long, but the real issue is both Intel and Nvidia have more compelling products on the shelves already.... and as we all know AMD are bleeding money as it is.

Still, if AMD pull a blinder and both are superb products, then they can of course turn things around.

 

[buggalugs]

Good News. AMD should have dumped TSMC ages ago.

 

[TheGuruStud]

All hail the Manufacturing God. The one true God. He really does care about us!

 

[Ebo]

Im just waiting for "greenland" to hit the market then its time to upgrade.

 

[midnightoil]

Uhm, dude, I was not wrong, this is from about 4-5 years ago and it was announced at a GloFo event. It was before 28nm was mainstream. No-one licensed nothing, it was a shared platform for IC development that could be used to produce the chips with any of the four companies and it was a big deal, as this had never been done before. Looks like Intel was involved as well. It was/is called the Common Platform Alliance.

You were and are wrong. Intel and TSMC were never involved in any kind of agreement or development ... perhaps they had representatives on a round table, but they were never ever involved beyond that, if they even went that far.

There were 4 major partners, and a host of minor parties. IBM, GF, Samsung, ST-M. IBM are out of the business and their foundry (mainly serving custom / semi-custom workstation chips) is owned by GF these days. ST are out entirely (though they continue to develop FD-SOI and license it to others).

Samsung and GF remain.

 

[HumanSmoke]

Still have to tape out two different chips, and as tests have proved, they don't share the same electrical properties (the A9 chips). I don't think there's anything easy or cheap about it .....

It also isn't that complicated. Aside from an interchange of cell libraries, compiler, and a second set of masks what else is involved. Proof please rather than hand-waving.

Also, GF 14nm FF LPP isn't copy exact .. it was originally supposed to be but they used different tools to save money ... hence delays and probably yield / volume / cost issues which got AMD to finally make the leap.

Again, please provide proof. GloFo's own announcement tends to indicate that the entire process, tool chain, test & validation will be mirrored by GloFo with the only real difference being the installed (non-litho) tool base

Through a proven level of fab synchronization never previously achieved outside of a single company, Samsung and GLOBALFOUNDRIES will use a coordinated copy-smart approach involving materials, process recipes, integration and tools.

and TSMC's FF+ process will have very, very limited / contested capacity at first, unlike Samsung LPP. Samsung's high power (and refined versions of LPE / LPP) come at the very end of '16, though I'd guess the LPE / LPP revisions will take priority over the HP version as AMD have gone with LP designs.

Yep, I'm going to ask for proof here as well. TSMC has been shipping series production (as opposed to risk) silicon on 16nmFF+ for some months.

TSMC 16nm FF+ and Samsung / GF 14nm FF LPP are completely different. FF+ is a high power process. LPP is a low power (but not SLP) process.

Hopefully LPP offers more against FF+ than 14nm LPE does against 16nmFF. Results on a like for like comparison so far don't auger well for Samsung/GloFo.

Uhm, dude, I was not wrong, this is from about 4-5 years ago and it was announced at a GloFo event. It was before 28nm was mainstream. No-one licensed nothing, it was a shared platform for IC development that could be used to produce the chips with any of the four companies and it was a big deal, as this had never been done before. Looks like Intel was involved as well. It was/is called the Common Platform Alliance.

It started to disintegrate after STM decided to pull out and the members process cadence started evolving at different paces.

 

[Jism]

It might be the reason why stocks on AMD went straight up: https://www.google.com/finance?q=NYSE:AMD

 

[HumanSmoke]

It might be the reason why stocks on AMD went straight up: https://www.google.com/finance?q=NYSE:AMD

If it is then it is rather slow picking up the news.

While we were in Burlington we heard the AMD has committed to 14LPP giving GF a volume first source customer to drive yield learning - Scotten Jones, SemiWiki, 11th November

AMD giving GloFo fabrication contracts should be pro forma. AMD and GloFo have a fairly draconian (for AMD) wafer supply agreement. Basically, a minimum purchase agreement whereby AMD pays a penalty if it does not fulfill its purchase quotas. AMD were in a no-win situation - order the silicon and have potential excess inventory/write-down issues, or be more conservative and incur penalties from GloFo. The agreement was expected to be revised by the end of the year - I'm guessing that the latest round of announcements have something to do with that

Lastly, as mentioned on our last quarter's earnings conference call, we are actively working with GlobalFoundries to re-profile our 2015 wafer commitments, in line with product demand in the fourth quarter of 2015 and into 2016. As of the end of the third quarter of 2015, we had purchases amounting to $631 million under the fifth amendment of the WSA. We anticipate concluding our wafer purchase reprofiling discussions with GlobalFoundries before the end of the year. - from AMD's Q3 earnings call

 

[btarunr]

What I'm wondering more is why AMD ditched their mobile GPU division!? Almost every phone has Adreno GPU now. Wouldn't that be like a good market for AMD if they kept it?

AMD sold its mobile GPU IP to Qualcomm years ago.

 

[HumanSmoke]

AMD sold its mobile GPU IP to Qualcomm years ago.

Just as the write-downs from the ATI acquisition started to bite. $65m for a mobile empire seems like wrist-cutting on a grand scale in retrospect, but AMD couldn't afford the R&D to develop mobile devices (or keep their set-top business which they offloaded to Broadcom) once the R&D budget was prioritized for greatest need.

 

[Jism]

Major brands are pushing for more profits by simply building their own in-house GPU.

 

[RejZoR]

Actually Qualcomm bought Adreno tech from AMD and you can see Adreno is just an anagram of Radeon. (same letters arranged differently)

I know that. What I don't understand is why AMD sold division for which devices get sold in millions and millions of units...

 

[uuuaaaaaa]

So a Transition for AMD into Samsung hands, hmm by 2017 will it be Samsung vs Intel and Samsung vs NV?

If I'm not mistaken that would mean the agreements that AMD has with Intel for x86 would not be valid, therefore no x86 cpu's could be produced...

 

[HumanSmoke]

I know that. What I don't understand is why AMD sold division for which devices get sold in millions and millions of units...

Dirk Meyer. AMD fighting on too many fronts with too little R&D.

Dirk decided that AMD wouldn't develop a strategy for the handheld GPU ( smartphones mainly) because he was pursuing high margin areas - basically servers (which is where his expertise lie). Dirk and the board gambled on Bulldozer hitting its targets to save the company rather than invest in low margin GPU IP for phones and other peripheral business (as Dirk saw it).

If I'm not mistaken that would mean the agreements that AMD has with Intel for x86 would not be valid, therefore no x86 cpu's could be produced...

Not necessarily. The cross-license agreement has a year grace period built in to allow a new owner to renegotiate. Not that it matters. Samsung is fully invested in ARM and I would doubt it has any use for x86 in its current strategy.

 

[TheLostSwede]

You were and are wrong. Intel and TSMC were never involved in any kind of agreement or development ... perhaps they had representatives on a round table, but they were never ever involved beyond that, if they even went that far.

There were 4 major partners, and a host of minor parties. IBM, GF, Samsung, ST-M. IBM are out of the business and their foundry (mainly serving custom / semi-custom workstation chips) is owned by GF these days. ST are out entirely (though they continue to develop FD-SOI and license it to others).

Samsung and GF remain.

Seriously? Do you get off on this shit?
This is the article I wrote about it in 2010 http://semiaccurate.com/2010/10/14/common-platform-alliance-key-foundry-success/

So Intel and TSMC might not have been part of the Common Platform Alliance, but they did work with GloFo, Samsung and IBM to work towards more common platforms http://hothardware.com/news/global-...ies-tsmc-and-samsung-announce-new-partnership

 

[Assimilator]

So GloFo is out in the cold, except for the mandatory agreement that's like an albatross around AMD's neck. If this Samsung partnership is successful it would make sense for AMD to move all their manufacturing (including next-gen console chips) there, which would leave GloFo with only Qualcomm as an anchor client - and fabbing ARM chips for smartphones isn't really something to be proud of.

As for Samsung 14nm vs TSMC 16nm... a ~3.3 billion transistor A9 is a far cry from a ~17 billion transistor Pascal (+ HBM). I'm sure Zen and Arctic Islands will be decent, but unless TSMC has really slipped up, Pascal will be better than decent - and AMD needs the latter, not the former, if its CPU business is to survive.

 

[efikkan]

Well, mid range and high end "Arctic Islands" GPUs are aleady taped out on TSMC's 16 nm process almost half a year ago. Samsung's 14 nm low-power Plus will not be suited for high end GPUs.

 

[Prima.Vera]

Hypothetically speaking, if Sammy buys AMD, will they also receive the x86 license??

 

[Steevo]

Well, mid range and high end "Arctic Islands" GPUs are aleady taped out on TSMC's 16 nm process almost half a year ago. Samsung's 14 nm low-power Plus will not be suited for high end GPUs.

So the fact that a 70% power and thermal reduction with the 14nm process and HBM means a todays card of 200W would draw a puny 70-80W at the same performance level means......

The scaling ability of newer low nm processes have been as limited by the memory interface as anything else, thus driving the need for lower voltage memory at the same and or higher performance levels...... why you ask? ODT On Die Termination, when the memory controller moved onto the silicon itself, the memory voltage has to be dealt with and stepped down to core voltage, and that is the job of resistors built into the die, and the larger amount of memory the more currant flow to maintain signal integrity, the more currant to terminate to prevent capacitive induction, the more heat....

Same goes for the number PCIe interface lanes built into the GPU core, so the ability to have a smaller node counts as much for the actual transistors doing computation as it does for the interfaces and power consumption for them.

I do imagine that AMD has a series of cards built on the LPP method that is pushing its limits, and possibly even why the release date is later on them, to work on power inside the core, as it's limited by the architecture as much as the process.

 

[medi01]

Difficult to tell, especially as Samsung (and GF's copy not quite exact) 14nm FF LPP process is a low power process and TSMC's 16nm FF+ is a high power process. But we do know that Samsung's LPE and LPP had vastly lower capacitance and leakage than Intel's (physically smaller) 14nm FF versions used in their mobile chips, also most probably much, much higher yields and much lower cost. I'd guess the same is also true of Samsung vs TSMC, but who knows.

Capacity won't be even remotely comparable. TSMC are very constrained for 16nm FF+. If AMD can get enough HBM2 and interposers, supply of Arctic Islands should be great (Zen too) ... Pascal may be little more than a dribble.

I was hoping AMD would switch to Samsung as their main foundry partner moving forward, and GF secondary, but didn't trust them to do so. My confidence in the new management has improved significantly now.

Wasn't there Samsung debacle with iPhone chips though? I.e. that Samsung 14nm chips were consuming more than TSMC 16nm?