Pity they don't want/can't use GF's 12nm for much more CPUs and GPUs this period of time.

The move to extend the WSA indicates that the company may continue to use 12 nm-class I/O dies in its processors for the foreseeable future.

Seems like this is a bit of speculation, with the knowing exactly whats been agreed on. Is GF able to do 10nm with confidence after the botched job on Intel? This reads like AMD has doubts about that too.

Complete speculation on my part, but maybe they will source SRAM for vcache with GF. SRAM does not scale well with node improvements. Just a pure guess though.

It would make sense for AMD to eventually transition these components to Samsung for 8nm eventually. Possibly TSMC at 7nm/6nm, but only if it actually makes since in terms of cost and making room to add additional chiplets perhaps by reducing the space and size these other components occupy.

I would take a 12nm GPU and CPU that was meant for 7nm, it would probably clock as well and just use 15 percent more power

SteevoI would take a 12nm GPU and CPU that was meant for 7nm, it would probably clock as well and just use 15 percent more power

It could be a great solution to improve the supply situation if they backported some of the newer designs to more available nodes. Obviously you wouldn't get the same performance from a rx6900xt but for (what used to be) entry level stuff it wouldn't make much of a difference and could bring prices down...

... and their margins down as well so won't happen :(

They should still be making 2nd Gen Threadrippers using GloFo too btw.. ;-)

looking back......GlobalFoundries should have spent the money on making 7nm. I guess no one expected the pandemic and the increase in demand for microchips

Investing in the divested.

Delicious.

ReallyBigMistakelooking back......GlobalFoundries should have spent the money on making 7nm. I guess no one expected the pandemic and the increase in demand for microchips

I second that.

DeathtoGnomesIs GF able to do 10nm with confidence after the botched job on Intel?

GlobalFoundries 7LP was fine, however GlobalFoundries didn't have investors nor any income to support volume ramp of 7LP. Initial capacity was 15,000 wspm for both 7LP and 7LP+EUV. To get more capacity the second phase of Malta needed to be built.

ReallyBigMistakeGlobalFoundries should have spent the money on making 7nm.

They did spend money on making 7nm. However, customers weren't willing to pay for capacity.

AMD #1 7LP customer opted to move to TSMC for first-source.
IBM #2 7LP customer opted to move to Samsung for first-source.
#3, #4, #5 did not need 7LP as soon or at all.

DeathtoGnomesThis reads like AMD has doubts about that too.

Fab8 is ramping up 45nm SOI; 45RF, 45RFe, 45CLO, 45SPCLO, 45SG01, 45SP3
Analysts who follow GloFo's node introductions also noted: 28SLP/28ULP and 22FDX/12FDX being integrated at Fab8, pre-ramp stage.

Going forward GlobalFoundries is less likely to maintain 14LPP/12LP/12LP+ supply with higher demand from 45nm, 28nm, 22nm, 12FDX platforms.
May 2021 => Removal of the exclusivity agreement of 14nm/12nm products.
December 2021 => Reduction of prices of 14nm/12nm wafers.

Why wasn't AMD willing to pay for GloFo's 7nm capacity? :kookoo::banghead: :confused:

AMD sabotages its own products by staying for too long on the old process nodes. It needs to decrease the power consumption and die sizes of those chipset thingies.

ARFWhy wasn't AMD willing to pay for GloFo's 7nm capacity? :kookoo::banghead: :confused:

AMD sabotages its own products by staying for too long on the old process nodes. It needs to decrease the power consumption and die sizes of those chipset thingies.

You do realise GF gave up on 7Nm, they're now specialists in Soic and other RF technology.
That's why they're used for the Io dies not core's.
Glofos 7Nm doesn't exist, anymore.

I know what happened. The question is why AMD continues to work with GloFo, and instead doesn't jump onto the new TSMC N3 and N5 process nodes.

ARFI know what happened. The question is why AMD continues to work with GloFo, and instead doesn't jump onto the new TSMC N3 and N5 process nodes.

Well contract's are going to be one point, they're obligated to and have been for years.
Constrained production capacity elsewhere likely isn't helping too.
And swapping to these lower nodes is not easy, it's usually a progressive process.

And when it comes to IO dies AMD can't afford to have it fd up, they need consistently made large batches usually best suited to very well established processes, without Io dies they're server lines fold, so do consumer lines, a badly yielding node doesn't typically mean the company manufacturing the wafer gets less money either so 3Nm low yields are shouldered by the buyer not the foundries.

ARFWhy wasn't AMD willing to pay for GloFo's 7nm capacity?

GlobalFoundries for 2018+ only had 15,000 wafers that had an asking price of $4 billion.
2020+ needed a second fab(and third for 5LP/3LP) which was >$15 billion(>$20 billion).

Net income for AMD(millions):

2020	$2,490
2019	$341
2018	$337
2017	$-33
2016	$-498
2015	$-660
2014	$-403
2013	$-83

Where would AMD pay for 7nm capacity? Being largely the sole customer of GloFo 7nm. Where as TSMC had customers, high net income, low-risk CapEx for expansion, etc. AMD didn't need to pay for capacity or a foundry for 7nm onwards with TSMC. The whole point of separating from GloFo to begin with was to NOT pay for foundry or capacity.

ARFThe question is why AMD continues to work with GloFo, and instead doesn't jump onto the new TSMC N3 and N5 process nodes.

GlobalFoundries-AMD has a WSA which obligates AMD to use GlobalFoundries. These more recent WSA's 2019/May 2021/December 2021 are all about minimizing costs w/ GloFo.

AMD being 28% of GlobalFoundries revenue in 2019 and 21% in 2020.
~$1.51816B for 2019; $5.422B with negative net income
~$1.18881B for 2020; $5.661B with negative net income
With 2021 being expected to be less than ~20%; Q3 2021 first positive net income for GlobalFoundries(primarily from older nodes than 14nm/12nm FinFETs)

Malta/Fab8 from the beginning was operated as a loss. Hence, why GlobalFoundries never made positive net income up till this point.
-0.604B in 2012
-0.876B in 2013
-1.5B in 2014
-1.3B in 2015
-0.426B in 2016

GloFo for AMD, basically crippled themselves, never actually being able to self-fund like TSMC, Samsung, Intel, UMC, SMIC. Mubadala stopped major funding GloFo sometime before the pivot.

I wonder if AMD intends to use GloFo for sectors that don't need bleeding or leading edge tech, like automotive, semi-custom solutions (moreso once they get Xilinx IP fully integrated), and certain cheap OEM components (like continued production of older Zen CPUs and GPUs for OEM-only markets). Those only need a relatively modern and mature process node for most of those tasks, and with tech becoming ever more integrated into vehicles (extreme cases like Tesla using an Nvidia or Radeon GPU aside), it'd be another ideal market for AMD's dedicated semi-custom division and with full knowledge of leveraging GloFo's 12/14nm process.

That aside, the other possibility is AMD looking into lucrative government (military) contracts. After all, Fab 8 only recently (in February 2021) got early certifications on the road to becoming a Trusted Foundry for the US government, and assuming all goes well, that certification may come late 2022 or 2023. That would also give AMD, the primary user of GloFo's 12/14nm process at Fab 8, plenty of leverage to use the reserved production lines for military hardware that don't need hardening at the chip level but need something relatively fast and modern. For GloFo, it's also a possible lead-in to getting additional funding via the CHIPS Act, and maybe a way to get some government sponsored funding for newer-gen node upgrades (whether it's restarting their 7nm equivalent or licensing Samsung's nodes again like they did with 14nm).

TechLurkerThat aside, the other possibility is AMD looking into lucrative government (military) contracts. After all, Fab 8 only recently (in February 2021) got early certifications on the road to becoming a Trusted Foundry for the US government, and assuming all goes well, that certification may come late 2022 or 2023.

In context, the Fab8 government deal is actually based off the 45nm SOI node:
gf.com/press-release/us-department-defense-partners-globalfoundries-manufacture-secure-chips-fab-8-upstate

Under the agreement, GF will provide a supply of chips built at Fab 8 on its differentiated 45nm SOI platform. The agreement is made possible by Fab 8’s compliance with U.S. International Traffic in Arms Regulations (ITAR) and highly restrictive Export Control Classification Numbers under the Export Administration Regulations (EAR).

Which extends to:
www.baesystems.com/en-us/article/bae-systems-collaborates-with-globalfoundries-to-produce-radiation-hardened-single-board-computers-for-space

The RAD510 SoCs are manufactured at GF’s Fab 10 facility in East Fishkill, N.Y., also a U.S. Department of Defense Category 1A Microelectronics Trusted Source. GF and BAE Systems have initiated the process of transitioning the manufacturing of 45 nm SOI technology to GF’s most advanced facility, Fab 8 in Malta, N.Y., making it available to the space community in the future.

Supply chain for Fab10 -> Fab8 of SOI nodes:
www.sec.gov/Archives/edgar/data/0001709048/000119312521290644/d192411dex104.htm
gf.com/press-release/globalfoundries-and-globalwafers-partnering-expand-semiconductor-wafer-supply

It is more likely any deal for GlobalFoundries is actually providing a roadmap for the SOI nodes;
www.skywatertechnology.com/press-releases/globalfoundries-and-skywater-technology-sign-mou-for-technology-development-to-strengthen-domestic-supply-assurance-for-u-s-government/

GF’s proven track record as a longtime supplier of secure semiconductor solutions for the defense industry dovetails with SkyWater’s focus on developing capabilities for enabling critical microelectronics technologies for national security applications. Notably, SkyWater is bringing to market a 90nm Strategic Rad-Hard process technology via an up to $170M DoD investment which, when combined with diverse GF platforms, will provide a richer portfolio of key defense technology offerings. The collaboration is also exploring possibilities for dual fab processing that would streamline access to secure technologies for defense programs. Additionally, as a part of the MOU, GF and SkyWater will align technology roadmaps and leverage unique and complementary capabilities for development and high-volume production path to market for advanced computing, artificial intelligence and related technologies.

RH90 is a fully depleted silicon-on-insulator complementary metal-oxide-semiconductor process specifically developed to produce electronics which can withstand harsh radiation environments. SkyWater aims to accelerate delivery of its offering to address U.S. government needs.

Converting Malta/Fab8 from 14LPP/12LP/12LP+ to 90nm RH-FDSOI -> 45nm RH-FDSOI -> 22nm RH-FDSOI -> 12nm RH-FDSOI is inline with GlobalFoundries prior obligations.

-----

TechLurkerThat would also give AMD, the primary user of GloFo's 12/14nm process at Fab 8, plenty of leverage to use the reserved production lines for military hardware that don't need hardening at the chip level but need something relatively fast and modern. For GloFo, it's also a possible lead-in to getting additional funding via the CHIPS Act, and maybe a way to get some government sponsored funding for newer-gen node upgrades (whether it's restarting their 7nm equivalent or licensing Samsung's nodes again like they did with 14nm).

The CHIPS Funding will most likely be used 12nm/sub-10nm FDSOI;
www.sitrigroup.com/wp-content/uploads/2019/11/FDSOI_RRAM-and-3D-are-key-technologies-for-AI-chips-and-5G_Jean-Rene-Lequepeys-from-Leti.pdf
www.networldeurope.eu/wp-content/uploads/2019/10/carloreita_2-10-2019-dresden.pdf

Technology solutions exist for 12nm and even sub-10nm nodes: no show-stopper identified at the device level

semiwiki.com/semiconductor-manufacturers/samsung-foundry/6913-semicon-west-the-fdsoi-ecosystem/

CEA Leti has modeled a 10nm FDSOI process and run test devices that match the modeled results. CEA Leti has also modeled 7nm and because 10nm did not need all of the performance boosters that are available Marie Semeria said she is confident 7nm is possible.

soiconsortium.eu/wp-content/uploads/2018/08/FDSOI-LETI_Shanghai_092018_V2.pdf
Page 10 minimum spec is 7nm node generation. GlobalFoundries has Monolithic 3D/CoolCube 3D for continuous scaling for example: 2-layer for 4nm/3nm FDSOI and 4-layer for 2nm/1.8nm FDSOI platforms.

Going forward the strategy for AMD is to minimize damage if GlobalFoundries drops 14nm/12nm FinFETs for aggressive FDSOI positioning.


====
Dresden = Already FDSOI
Malta = Transitioning to FDSOI
Singapore = Fab7H/"New Fab Singapore" is a replacement for Fab11P1(130/180nm)/P2(22FDX)(Chengdu)
gf.com/press-release/globalfoundries-breaks-ground-new-fab-singapore
$10B(undisclosed) with half-being paid by GloFo(>$4B) and half-being paid by Customers/Government(>$4B from Singapore Economic Development Board and with co-investments from committed customers)
Which is similar price arrangement as Chengdu; ~$5B from GloFo and ~$5B from MediaTek(/AMD China/Rockchip)/Chengdu Municipal Gov.

For example of internal dual-source reserving:
www.sec.gov/Archives/edgar/data/772406/000077240621000030/ex105gfcapacityagreement.htm
GF shall have the exclusive right to decide at which of Fab 1 or Fab 7 (or both) a new Product will be taped-out, qualified, and manufactured. => If production will be in both Fabs, Customer and GF will make commercially reasonable efforts to qualify the Product at both Fabs.

An oldie:

AMD Foundry Dresden and Chartered Singapore => Now, GlobalFoundries Fab1/Fab7.

On another note, AMD disclosed they lost manufacturing capability of 32SHP/28SHP/28A at GlobalFoundries sometime in 2020 for 32SHP and 2021 for 28SHP/28A.

I believe this also covers the ~360 mm2 A9-9820 Cato chips.

2x4-core + RX 350(iGPU) + Quad-channel DDR3

In relation to volume of wafers bought, 2021 onwards it is only 14nm/12nm products as far as official already out products go.

Any Malta pivots away from 14nm/12nm FinFETs by GloFo will be disclosed next year.
Samsung-GlobalFoundries => April 2014 // Killed 20LPM/20SHP/20AN/14XM
GlobalFoundries Ends 7nm => August 2018 // Killed 7LP/5LP/3LP
Potential date for FinFETs being killed off => April 2022-August 2022 // Killing production of 14LPP/12LP/12LP+ to fit in 22FDX(Co-fabbed w/ Singapore+Dresden)/12FDX(Co-Fabbed w/ Dresden).

Getting away from AMD and looking at the American picture:
Intel has 22FFL and 14GP at Oregon, currently there is no 22nm at Malta and 14LPP/12LP/12LP+ does not compete against Intel's FinFETs.

The MOSIS Service, is collaborating with Intel Corporation to provide access to 22nm FinFET Low-power (22FFL) process technology for the microelectronics design community through the Intel Custom Foundry. To encourage participation in the MOSIS Service offerings of Intel’s multi-project wafer (MPW) fabrication runs, the Office of the Under Secretary of Defense for Research and Engineering (OUSD(R&E)) Trusted and Assured Microelectronics (T&AM) Program aims to potentially sponsor MPW runs in government fiscal year (GFY)-2020 and GFY-2021.

W/ PG's Gov involvement leading towards Gov:14LPP/12LP/12LP+ designs being shifted towards Gov:14GP instead. Which means going forward GloFo is less likely to win out trying to use FinFETs on CHIPS. When Intel is better at service and end quality, a.k.a. money well spent.

Samsung has 17LPV/18FDS and 14LPE/14LPP/14LPC/14LPU/11LPP at Austin, there is no response to 17LPV(20nm FinFET<90CPP> w/ 28nm BEOL)/18FDS(20nm FDSOI<90CPP> w/ 14nm BEOL) and Samsung has a second source in Korea for FinFETs/FDSOI.


New Capacity/Fab growth:
Intel Oregon/Arizona FinFETs
TSMC Arizona FinFETs
Samsung Texas (Majority FinFETs for Perf/Minority FDSOI for Power)

GlobalFoundries is the only foundry that has FDSOI for Performance and they don't even fab it in America.