Wafer Supply Agreement Discussion

[NostaSeronx]

Wafer Supply Agreement;
http://www.sec.gov/Archives/edgar/data/2488/000119312509046101/dex105.htm

Amendment #1;
http://www.sec.gov/Archives/edgar/data/2488/000119312511200905/dex101.htm

Amendment #2;
http://www.sec.gov/Archives/edgar/data/2488/000119312512222047/d312884dex101.htm

Amendment #3;
http://www.sec.gov/Archives/edgar/data/2488/000119312513069422/d486815dex1034c.htm

Amendment #4;
http://www.sec.gov/Archives/edgar/data/2488/000119312514177314/d714367dex101.htm

@mrmt, I found the and in Amend#4;

The purchase price for the 2014 MPU and GPU Production Wafers shall be paid in accordance with the payment provisions set forth in the Agreement, subject to the following modifications: (X) FoundryCo shall accept [****]-day payment terms from January 1, 2014 through [****], 2014

 

[jpiniero]

I like how AMD is required to wire the money to GloFo at a bank on the Cayman Islands like they are some kind of drug lord.

 

[NostaSeronx]

I think Amendment #3 is the most important one. It essentially removed the exclusive node requirement for any standard node. AMD being forced to pay for R&D and usage of the SHP nodes are over.

The issue is that there are many standard nodes now.

1st Generation 28-nm LP - SLP
2nd Generation 28-nm LP - HPP
1st Generation 28-nm ETSOI - ???
2nd Generation 28-nm ETSOI - ???
1st Generation 20-nm HP - LPM
1st Generation FinFET - LPE
2nd Generation FinFET - LPP
1st Generation 14-nm ETSOI - ???

Fab 1 will only be doing ETSOI. (Other than existing 28-nm SHP orders)
Fab 8 will be doing FinFETs and ETSOI.

 

[positivedoppler]

what is carrizo going to be on?

 

[mrmt]

what is carrizo going to be on?

Vanilla 28nm bulk.

 

[NostaSeronx]

what is carrizo going to be on?

For the 28-nm Technology Node for 2014/2015 MPU and GPU Products.

It can be any of these three;
28-nm HPP
28-nm ETSOI (80% 2nd Generation / 20% 1st Generation)
28-nm LPH

HPP and LPH perform parametrically worse than 28-nm HP/HPM at TSMC. So, if it is these two nodes it is most likely going to get waived. Going on anything but ETSOI will be a bad idea economically and competitively.

I have not seen any support for DDR4 with 28-nm HPP/LPH. We know for sure about 28-nm FDSOI/ETSOI; http://www.cadence.com/cadence/newsroom/press_releases/pages/pr.aspx?xml=051514_DDR4ST
Nevermind for above; http://i.imgur.com/jJqjRdD.png

http://i.imgur.com/5gh9q13.jpg <-- 28-nm FDSOI 1st Generation versus 28-nm HPP.

Then, you have;

A second generation of 28nm FD-SOI, which implements source/drain engineering, delivers a 25% performance gain over the high-end 28nm high-K/metal gate (HKMG) bulk with slightly lower process wafer cost, or achieve the same performance as 20nm bulk but at the cost of 28nm HKMG. This new generation uses in-situ boron doped SiGe epi for PMOS and phosphorus or arsenide doped Si or Si(C) Epi for NMOS, as well as forward back bias. Note that this is source/drain engineering &#8211; the channel remains undoped, so the advantage of FD-SOI in terms of avoiding RDF is maintained.

2nd Generation other than density concerns;
http://i.imgur.com/0Ldzbuc.jpg

 

[positivedoppler]

do you have any source that show GF has soi?

 

[positivedoppler]

Vanilla 28nm bulk.

Is there a source for that?

 

[NostaSeronx]

do you have any source that show GF has soi?

From July 10, 2014;

GlobalFoundries committed to using ST&#8217;s FD-SOI tech for 28nm and 20nm production in 2012, and expects to put it into volume production by the end of this year, for 28nm and 14nm processes. In May, STMicroelectronics announced that Samsung would use ST&#8217;s 28nm FD-SOI tech for foundry customers. Samsung plans to offer the process in early 2015.

From October 17, 2014;

Next up, Manfred Horstmann, Director of Products & Integration for GlobalFoundries in Dresden said that FD-SOI would be their focus for the next few years. They&#8217;re also calling it ET-SOI (for extremely thin), and he said it&#8217;s the right solution for SOCs, especially with back biasing. Plus, it&#8217;s good for the fab because they can leverage their existing tool park. Asked if they were seeing interest, he said yes. Asked if they have customers lined up, he said yes. So watch this space &#8211; there&#8217;ll be news soon!

Is there a source for that?

GF28A denotes to: "28nm Global Foundries bulk CMOS process"

It is a misnomer as GF28A is not GF28 as GF28 denotes to "28nm Global Foundries bulk CMOS process."

GF28 != GF28A and resumes can be wrong. The position of G28A and GF20A/GF20AN is adequate to 28FD and 20FD/14FD.

 

[positivedoppler]

thanks for the insight Nosta. Do you know if its easy to port an existing 28nm design to soi? If so, why doesnt amd port all their r2 graphics over to gf soi?

 

[mrmt]

thanks for the insight Nosta. Do you know if its easy to port an existing 28nm design to soi? If so, why doesnt amd port all their r2 graphics over to gf soi?

Carrizo will be 28nm Bulk because there is no 20nm working node at Globalfoundries. They screwed up both its 20nm offers and its "14nm"XM. That's why they had to go to Samsung to license a node for 14nm.

As for SOI, AMD ditched SOI in 2012, it was one of the first things Rory Read did at AMD, along with killing the CMT developments. AMD shipment volumes are too small, only 13MM chips in bleeding edge node (GPU + CPU, excluding consoles) and those are projected to fall next year. They have no room for fancy experiments like SOI, especially because they have no room on the kind of market they are trying to reach. AMD will stick to bulk, standard nodes for the foreseeable future.

If you net the console sales, AMD entire GPU + CPU business is shipping less units than Intel is shipping on the tablet market.

 

[NostaSeronx]

Do you know if its easy to port an existing 28nm design to soi?

I don't know about easy. It should be;
1. Straightforward (Bulk -> FDSOI)
2. Similar (Same tools)

If so, why doesnt amd port all their r2 graphics over to gf soi?

The reason is timing FDSOI is only hitting production now.

As for SOI, AMD ditched SOI in 2012, it was one of the first things Rory Read did at AMD

http://www.soiconsortium.org/news-e...ation-mobile-and-consumer-applications-30.php

GlobalFoundries ditched PDSOI in 2011 for FDSOI.

From ChipHell in September 2011;

http://www.brightsideofnews.com/2011/08/30/globalfoundries-announces-20nm-process-no-to-soi/
http://www.brightsideofnews.com/2011/09/11/rumors-14nm-node-and-450mm-wafers-by-2015/

Best way to train AMD for FDSOI? Force them on bulk.

Two more bits for WSA talk.

The price for GPU products will be determined by the parties when The Foundry Company is able to begin manufacturing GPU products for us.

The Foundry Company will provide certain foundry services to us. In August, 2008, we, on behalf of The Foundry Company, extended our JDA with IBM through December 31, 2015 and added 32-nanometer and 22-nanometer &#8220;bulk&#8221; Industry Standard technology projects to the existing &#8220;SOI&#8221; High-Performance development and research projects as part of the strategy to develop the manufacturing technology for a broader potential Foundry Company customer base. The JDA will be assigned to The Foundry Company upon consummation of The Foundry Company manufacturing joint venture transaction.

In addition, once The Foundry Company establishes a 32-nanometer qualified process, we would purchase from The Foundry Company, where competitive, specified percentages of our GPU requirements, which percentage is expected to increase over a five-year period. We agreed not to sell, transfer or dispose of all or substantially all of our assets related to GPU products and related technology to any third party without The Foundry Company&#8217;s consent, unless the transferee agrees to be bound by the terms of the Wafer Supply Agreement, including its minimum purchase obligations, where competitive, with respect to GPU products.

That explains why Northern Islands was 32nm and Sea Islands showed 22nm.
http://semiaccurate.com/2013/05/20/...-red-flags-for-the-company/amd_roadmap_tiran/

For those participating for the same above bulk nodes;

32nm Bulk-Industry Standard Semiconductor Process Technology
1. Infineon Technologies A.G.
2. Samsung Electronics Co., Ltd
3. Toshiba Corporation
4. ST Microelectronics N.V.
5. GLOBALFOUNDRIES, Inc
6. NEC Electronics Corporation
22nm Bulk-Industry Standard Semiconductor Process Technology
1. ST Microelectronics N.V.
2. Samsung Electronics Co., Ltd
3. GLOBALFOUNDRIES, Inc

All three in the 22/20-nm alliance are now vested in FDSOI/ETSOI. While, only Samsung is paying for FinFETs between GlobalFoundries/Samsung.

Just for a fun off-topic bit in the same timezone Intel;

The following is a list of our main microprocessor competitors by market segment:​

&#8226; Notebook: AMD and VIA
&#8226; Netbook: AMD, NVIDIA, QUALCOMM, and VIA
&#8226; Desktop: AMD and VIA
&#8226; Server/Workstation: AMD, IBM, and Sun Microsystems
&#8226; Embedded: AMD, Freescale Semiconductor, Inc., and Texas Instruments Incorporated
&#8226; Handheld: QUALCOMM

 

[erunion]

thanks for the insight Nosta. Do you know if its easy to port an existing 28nm design to soi? If so, why doesnt amd port all their r2 graphics over to gf soi?

not to burst your bubble, but this forum doesn't take nasto seriously. He makes bad assumptions off wrong internet rumors. His track record is available to see.

 

[positivedoppler]

not to burst your bubble, but this forum doesn't take nasto seriously. He makes bad assumptions off wrong internet rumors. His track record is available to see.

Soi is advertised on gf's website. It is believable. AMD advertise a 30% performance boost at 15 watts. It matches soi advertisement

 

[NostaSeronx]

Soi is advertised on gf's website. It is believable. AMD advertise a 30% performance boost at 15 watts. It matches soi advertisement

This key technology has proven it can deliver 30% higher speed at the same power and up to 50% greater power efficiency at the same performance as bulk processes—at comparable cost.

It does say that, but it is to convenient and doesn't match comparisons.

28-nm SHP GF = 28-nm HP TSMC, in performance, according to AMD Linkedin Profiles. 28-nm SHP and HP to 28-nm HPP/LPH would mean a slower APU/GPU.

 

[Idontcare]

thanks for the insight Nosta. Do you know if its easy to port an existing 28nm design to soi? If so, why doesnt amd port all their r2 graphics over to gf soi?

It is beyond "not easy" to port a bulk-Si design over to SOI, be it partially or fully depleted.

Porting an existing design from bulk-Si to SOI requires no less time, money, or engineering effort than porting said existing design to a different node entirely

(28nm bulk -> 28nm SOI entails the same engineering timeline and expense as 28nm bulk -> 20nm bulk or 20nm SOI)

If you look at history, the only companies who have elected to critically rely on SOI are companies that have been receding from the lime-light, not progressing towards it.

IBM's footing in the semiconductor world has only been eroding over the past decade, as has AMD's. They both did great things with SOI, but in the end SOI did not alter their respective fates.

The companies who seized the day and grew like gangbusters (and still are) did so by avoiding SOI altogether.

There is a solidly good reason for that.

 

[NostaSeronx]

PDSOI design time = Long (Effects with FBE)
FinFET design time = Long (Capacitance)
Bulk design time = Medium (Fluctuations)
FDSOI and SOI FinFETs design time = Short and Medium

Those who will seize the future and grow like gangbusters will do so by going FD-2D or FD-3D SOI.

28-nm Si to 28-nm FDSOI is not as long as you think it is IDC. It is much, much shorter than that.

If a semiconductor made two teams one for 28-nm Bulk and one for 28-nm FDSOI. Then, made them build the same processor with the same tools. The FDSOI processor will finish first, and with extra time a 2nd generation could be planned. This 2nd generation would most likely better use of what is native to FDSOI.

With three SmartCut FD-2D/FD-3D wafer foundries; Shin-Etsu, SOITEC, and SunEdison.
With three EDAs supporting FDSOI; Synposis, Cadence, and Mentor.
With two foundries; Samsung and GlobalFoundries.

Synpase Design and Verisilicon Microelectronics, have made Bulk to FDSOI ports. It is easier for them, and they do see FDSOI to be more mainstream than FinFETs.

GlobalFoundries Fab 1 is exclusively after 28-nm SHP, FD-2D; for 28-nm, 14-nm, 10-nm, and FD-3D; for 14-nm(1st Gen), and 10-nm(2nd Gen). There is no bulk at Fab 1 other than 28-nm SHP.
GlobalFoundries Fab 8 is 75% FD-2D and 25% Samsung FinFET. Samsung's 10-nm FinFET is FD-3D, so we know where that is going.

It is quite clear that GlobalFoundries has a FDSOI bias. With that said GlobalFoundries will push AMD to ETSOI nodes. This will most likely planned way ahead of time.

SOI is no longer the wounded beast that was PDSOI. It's recovered with extra speed for less given power, and it's fully depleted.

====
Intel by externally pushing bulk as the node of choice, Intel has cemented their lead. No design on TSMC Bulk or Common Platform Bulk will be as efficient as Intel's Bulk. While PD-SOI did have issues, majority of the problems are gone with FD-SOI. ET-SOI from GlobalFoundries will make AMD chips; Faster and Cooler. While, also reducing the cost as 28-nm SHP is very expensive.

 

[Ajay]

Carrizo will be 28nm Bulk because there is no 20nm working node at Globalfoundries. They screwed up both its 20nm offers and its "14nm"XM. That's why they had to go to Samsung to license a node for 14nm.

At least GF has gotten serious about ramping up Malta by appointing a new GM: http://www.saratoga.com/saratogabus...ew-senior-vp-gm-at-globalfoundries-malta.html

Of course, they are going to be late on 20nm (compared to TSMC & Samsung) and Samsung is doing it's best to make sure their 14FF is ramped up before GF (as in, on purpose). I'm sure that Samsung wants the majority of Apple's business - even if Apple pushed for the second source arrangement.

So right now GF is stuck being a lagging node provider.

 

[NostaSeronx]

So right now GF is stuck being a lagging node provider.

GlobalFoundries is on time with 14-nm FD-SOI(STM), and 14-nm SOI FinFETs(IBM). Both are positioned against Intel's 14-nm FinFET process in time to volume. Decoded; When Intel's 14-nm FinFETs hit volume, so will 14-nm Planar FD-SOI and 14-nm SOI FinFETs. SOI FinFETs take up the PD-SOI mantle of high performance. FD-SOI replaces the bulk mantle of being mainstream.

It's only the bulk nodes that are delayed or canned. The WSA in regards to 20-nm bulk will be iffy at best.

Off-topic: Nvidia states POWER8+, but I only see POWER9;
http://www.linkedin.com/pub/samarth-mittal/17/172/446
http://www.linkedin.com/pub/byron-scott/5/45a/20b
http://www.linkedin.com/in/tvluong
http://www.linkedin.com/pub/hwajoon-oh/4b/407/11a
http://www.linkedin.com/pub/ivan-vo/89/944/b48
^-- This will only become on topic, when IBM sells their foundries to GlobalFoundries, and also sets up a WSA, imho.

 

[mrmt]

At least GF has gotten serious about ramping up Malta by appointing a new GM: http://www.saratoga.com/saratogabus...ew-senior-vp-gm-at-globalfoundries-malta.html

They always say they are serious about everything but never deliver. There seems to be a gap between what management think the company can do and what it can actually deliver. I hope they can go a bit further than powerpoint for the next nodes.

 

[Idontcare]

At least GF has gotten serious about ramping up Malta by appointing a new GM: http://www.saratoga.com/saratogabus...ew-senior-vp-gm-at-globalfoundries-malta.html

Of course, they are going to be late on 20nm (compared to TSMC & Samsung) and Samsung is doing it's best to make sure their 14FF is ramped up before GF (as in, on purpose). I'm sure that Samsung wants the majority of Apple's business - even if Apple pushed for the second source arrangement.

So right now GF is stuck being a lagging node provider.

GF is a broken foundry model (from a business perspective) because they don't have R&D for internal node development the likes of Samsung or TSMC.

And the gating problem for them is that it is a four-year process from start to finish to develop your own process node. It is not something that can be done on-the-fly with a moment's notice.

For all practical considerations, GF will always be dependent on someone else developing a node and licensing it to GF - be it IBM, Samsung, STMicro, etc. - and this will always come to be a factor in any fabless company's business decision matrix.

It isn't a position I'd wish upon any foundry, but it is the position they find themselves in, and are doing the best they can to make lemonade from those lemons. But it is a position that leaves them intrinsically and inherently weaker than their competitors (TSMC, Samsung, etc.).

Hence the necessity of having absurdly binding contracts with AMD as the WSA.

 

[III-V]

Didn't GF pick up IBM's node R&D? Or am I having a faulty memory?

 

[NostaSeronx]

I've been decoding the 4th Amendment to the WSA between AMD & GlobalFoundries.

2.5D-IC might be one of the things involved with it. Which might be the reason why we haven't seen the GPUs at GlobalFoundries, yet.

Didn't GF pick up IBM's node R&D? Or am I having a faulty memory?

IBM is buying their foundries to GlobalFoundries. While starting a relationship like CEA-Leti(Advanced R&D) and STMicroelectronics(Foundry).

IBM -> $2B + Foundries -> GlobalFoundries

Applied Micro has this to say;

The recent uncertainty regarding IBM&#8217;s foundry business and its possible sale to Global Foundries has impacted customer behavior, causing an acceleration of the transition away from the PowerPC architecture. This has resulted in the decline of this legacy business to be more front-end loaded in our fiscal year than originally expected.

Not related to above but is WSA-relative;

ATIC&#8217;s obligation to provide funding is subject to certain conditions, including, among other things, the accuracy, in all material respects, of GLOBALFOUNDRIES representations and warranties in the Funding Agreement, the absence of a material adverse effect of GLOBALFOUNDRIES, the absence of a material adverse effect on AMD that could reasonably be expected to materially and adversely affect AMD&#8217;s performance of its obligations under the Wafer Supply Agreement, and the absence of a material breach or default by GLOBALFOUNDRIES or AMD under the provisions of any document related to the transactions.

With respect to Phase II, ATIC&#8217;s obligation to provide funding is subject to certain additional conditions, including, among other things: (i) the continuing effectiveness of AMD&#8217;s IBM Participation Agreement; and (ii) the availability of New York and Dresden grants in amounts not materially different from the amounts contemplated in the five-year capital plan; (iii) AMD will have secured for GLOBALFOUNDRIES &#8220;AMD-specific Have Made&#8221; rights (defined as the right of AMD to have unlimited volumes of products, including microprocessors, made for AMD and its subsidiaries by GLOBALFOUNDRIES); (iv) GLOBALFOUNDRIES will have achieved targets for cumulative revenue and cumulative gross margin; and (v) GLOBALFOUNDRIES will have achieved certain strategic milestones relating to the groundbreaking and build out of the Abu Dhabi fab, progress on certain technologies and third party customer interest and revenue (&#8220;Phase II Funding Conditions&#8221.

 

[mrmt]

Didn't GF pick up IBM's node R&D? Or am I having a faulty memory?

No.

 

[mrmt]

GF is a broken foundry model (from a business perspective) because they don't have R&D for internal node development the likes of Samsung or TSMC.

Isn't that the same kind of limitation that UMC and SMIC have to deal?