icon-share-facebook Facebook
icon-share-twitter Twitter
icon-share-linkin Linkedin
CH
創業小聚
News
Executive interview Insights Interview with VCs International partnership Startups to watch Startups Energy Sports Web3 AI Communications Technology Hardware Cloud Semiconductor ESG AR/VR Lifestyle Travel Security Mobile Commerce Entertainment MarTech FinTech Edtech Healthcare EV Media BioTech Digital Content
Events
創業小聚(Chinese)
Follow
【VC Column】Capital is not the problem in building semiconductor fabs TSMC and AWS: Comparing Two Infrastructure Giants Why Russia can’t replace TSMC How AMD Left GlobalFoundries for TSMC India's Semiconductor Failure TSMC’s Crypto Customers
Insights

How AMD Left GlobalFoundries for TSMC

Asianometry 2022/01/26
 Asianometry
Asianometry

AMD began as an integrated design manufacturer competing directly with Intel - meaning they designed and made their own chips. When the company's internal manufacturing operations started to lag behind Intel, the company transitioned into being a fabless manufacturer.

Even after the transition, its foundry partner GlobalFoundries struggled to keep up with the company's technical and yield demands. Eventually the company would turn to the Taiwanese giant TSMC for its leading edge processes. Today, TSMC is AMD's closest foundry partner and the manufacturer of some of its best products.

In the video, we will look at the chip designer's move away from its manufacturing roots, its up-and-down relationship with GlobalFoundries, and how it came to fab its most advanced products at TSMC.

AMD's Manufacturing Struggles

By the late 2000s, AMD needed help. They turned losses for over two years starting in 2006. And the company's products were losing ground to Intel and the Core microprocessor architecture. This was despite a number of successful anti-trust suits against Intel targeting its anticompetitive practices. Intel continued to lead its smaller rival in technological prowess and product category dominance.

In 2007, AMD's first quad-core chip offering codenamed Barcelona lagged its Intel competitor to the market by nine months. The late release - and the product's subsequent tepid cost-performance offering - squandered an opportunity to grow its 20%-ish market share.

Intel seemed to be firing on all cylinders. It announced that all of its foundries had moved from a 65nm to 45nm process, a year ahead of AMD. Beyond that, Intel said that its 32nm and 22nm processes would roll out in their usual 2-year cadence. This model put AMD at even more of a disadvantage, as they had to keep up and leading-edge foundries were getting increasingly expensive.

As an IDM that did both the design and manufacturing, AMD needed to spend millions of dollars building fabs. It needed sales or sales commitments from potential customers to get that money. But those potential customers won’t make those commitments if they feel that the fab won’t be delivered on time.

The result was a chicken and egg dilemma that left AMD in a downward spiral. Even with generous tax breaks coming from Germany for building its Fabs 30 and 36 in Dresden, the company had to halt construction of those Dresden fabs as well as a few New York fabs to conserve cash.

In October 2008, AMD announced that they would be spinning their manufacturing fabs off into a separate company. Hector Ruiz, AMD's second CEO, exited after seven consecutive quarterly losses.

GlobalFoundries's Business Strategy

The spinoff would allow AMD to focus on being a fabless company and get a capital injection. The emirate of Abu Dhabi came in as that white-horse investor. They had already bought 8% of the company in 2007 as part of an earlier deal. Now they were moving to get more involved.

The spinoff would have Abu Dhabi paying $700 million for 55% of the new FoundryCo (later to be named GlobalFoundries) and another $300 million to raise their ownership share of AMD to 19.3%. The country continues to be a major shareholder in AMD, with 6.9% ownership share as of this writing.

The transaction was consummated in March 2009. AMD owned 34% share in GlobalFoundries and became its first customer. GlobalFoundries sought to become a leading player in the semiconductor space by consolidating other foundries into a single major player that can challenge the current market leaders.

In acquiring their customers, the idea was that the company could pool together enough capital to fund the research and development for future processes as well as the immense capital expenditures to build fabs.

Six months later, Abu Dhabi took a step towards executing on that strategy by purchasing struggling Singaporean player Chartered Semiconductor. I did a video about them previously that you should check out. The acquisition instantly made GlobalFoundries the third biggest foundry in the market. Things seemed well.

The Wafer Agreement

In March 2009, GlobalFoundries struck what is called a wafer supply agreement that dictated the terms of their relationship. The agreement remains in force to this day, binding the two companies together. But over the years, several significant amendments have been attached, progressively giving AMD the freedom to pursue business with TSMC.

It should be noted that by then AMD and TSMC already had a working relationship. In 2006, AMD purchased Canadian graphics chip maker ATI in a deal worth about $5.4 billion. ATI had been one of TSMC's biggest customers, and that relationship continued for a few years even after the merger.

So it would seem that AMD knows the benefits that going to TSMC might give them. But the 2009 Wafer Agreement mandated that AMD purchase all of its products from GlobalFoundries. The price at which these products would be sold was calculated on a cost-plus model, the cost of the goods along with a percentage markup. If GlobalFoundries was unable to meet AMD's volume demands though, AMD can bring in another foundry to serve as a second source for a time.

To keep TSMC from being brought in as a second source, GlobalFoundries had to capably bring its manufacturing processes to the leading edge, matching blows with Intel. The key issue would be the development and rollout of its next generation process - 32nm.

Falling Behind

They were failing at the task. In 2010, just about a year into the new relationship, GlobalFoundries found itself seven or eight quarters behind Intel. Santa Clara had already successfully transitioned their fabs to a 32nm process, with 28nm on track for 2012. GlobalFoundries was still working out the kinks for 32nm, and the delays were hitting AMD's product rollout.

In the 2Q 2010 earnings call, then-CEO Dirk Meyer said that due to "a slower than anticipated progress of [the] 32nm yield curve", AMD had to delay its quad-core 32nm Llano APU product a year to the second half of 2011. To fill the void, AMD brought up the product rollout of its dual-core APU product Ontario.

(FYI, an APU is an AMD marketing term for a system on a chip die that included both a CPU and GPU. At the time, they made up a significant percentage of AMD's sales.)

Ontario was fabbed at TSMC at 40nm. That node was a troubled one for TSMC. Yes, TSMC is capable of having issues too. Some reports claimed that the 40nm yields were as low as 20-30%. The fiasco basically cost TSMC’s then-CEO Rick Tsai his job, as he was reassigned and the octogenarian founder Morris Chang took over once more.

Despite AMD being affected by this - its Radeon series 5000 graphics cards were fabbed on 40nm - AMD still chose TSMC to make those Ontario chips. GlobalFoundries at the time was digesting its Chartered Semiconductors acquisition and did not have the capacity to handle the job. So TSMC took over an AMD CPU product for the first time, poured $1.6 billion to increase its factory capacity and Ontario shipped on time in January 2011.

Compared to how TSMC handled Ontario at 40nm, the GlobalFoundries 32nm delay was immensely frustrating and AMD sought to keep it from happening again. Thus in April 2011, a few months before the Llano APU launch, the two companies announced changes to the Wafer Supply Agreement.

This new agreement changed the way AMD paid GlobalFoundries for one year. Where before they used the aforementioned "cost plus" model to determine the wafer prices, this new agreement changes it to where AMD would only pay for "good die". So GlobalFoundries was incentivized to improve its yields otherwise it would be eating the cost for bad wafers. If the yields do improve then AMD would pay $400 million in bonus payments. Again this would only be in effect for one year - presumably by then GlobalFoundries would have their cards straight on 32nm.

It appears that the new wafer agreement cost GlobalFoundries quite dearly. In Q3 2011, the foundry fired their CEO Doug Grose as well as a bunch of other senior people. The company had eaten the costs for bad wafers all year long.

In 2012, just three years after the spinoff, AMD divested its final stake in GlobalFoundries, granting the company full economic independence. In the transaction, AMD gave GlobalFoundries $425 million and then amended its wafer agreement once more so to allow AMD to temporarily fab some of their 28nm APU products at TSMC. They never really gave a reason why. My bet is because of capacity issues again from the Chartered acquisition.

Intel by now was already heading to 22nm with Ivy Bridge. But things were starting to change. In 2014 AMD appointed a new CEO, Dr. Lisa Su, and she began her tenure by emphasizing the primacy of the technology. AMD needed to have chips superior to Intel. In order to do that, they needed access to the best process manufacturing technologies.

GlobalFoundries Exits the Leading Edge

In order to diffuse immense and continually rising node development costs, GlobalFoundries teamed up with South Korean giant Samsung Electronics. They announced on April 2014 that they would work together to bring out their 14nm process by the end of the year or later.

The collaboration with Samsung allowed GloFo to deliver their next generation 14nm node on time in the first half of 2015. And that got them the nod for AMD's Radeon RX 400 Series GPUs. Released in June 2016, it was a good product that won awards for its cost-performance ratio. And for what it is worth, GlobalFoundries' 14nm seemed to have yielded rather well for its customers.

But AMD wanted more access to leading edge nodes. So in August 2016 it went back to the Wafer Supply Agreement with GlobalFoundries and amended it once more, its Sixth Amendment to the Agreement.

It first established that GlobalFoundries would remain the preferred supplier for the next generation 7nm process. The two companies will collaborate on fleshing out that process and getting it to work.

But at the same time, the amendment would allow AMD the possibility of pursuing 14nm and 7nm process jobs with other foundries in exchange for financial considerations. In other words, AMD would have to pay GloFo for the privilege of not using their process.

It is likely that AMD did this in case they would need to secure second-sourcing capability for their ground-breaking Zen microarchitecture. At this point, AMD was basically betting the company on Zen succeeding. Failing because of a supply issue was unacceptable.

GloFo and Samsung continued their technology collaboration and that allowed GlobalFoundries to announce its 12nm process in 2017 - an iterative improvement on the 14nm. With these excellent, high quality nodes available to them, AMD would launch two generations of Zen to great success.

Despite this successful launch, however, GlobalFoundries was starting to come to terms with its fate. They had plenty of incentives for getting to 7nm, but the economics were getting increasingly worse. Going to 7nm means acquiring and mastering the next major paradigm in lithography: EUV.

For GlobalFoundries to get EUV to work, everything from the light source to the recipes had to change - a retooling job that would probably cost tens of billions of dollars. GloFo would have to find some way to get that kind of capital and its Middle East backers had finally had enough.

And on the other side, just how many customers really needed EUV 7nm? AMD would be the biggest, but other than that, you have just a few flagship big-market customers like Apple, Nvidia, Bitmain, and Qualcomm. Those are demanding, pain-in-the-butt customers and GlobalFoundries would have to compete with TSMC and Samsung for them.

Why put in the pain and effort when there is plenty of business in other processes?

So in August 2018, two years into developing 7nm, the company announced that it was putting development of that technology on hold. Instead, it would focus on those two 14nm and 12nm processes - perfectly good for most customers - as well as other speciality processes for niche markets.

TSMC 7nm

With GloFo exiting the 7nm race, AMD needed a new foundry for its upcoming Zen 2 microarchitectures. You see, this new Zen 2 product would be utilizing a brand new physical design paradigm as compared to previous Zen generations. They would be moving away from a single monolithic design to what is called a "chiplet" design tying together several smaller dies.

Smaller dies have better yields and heat distribution. It makes intuitive sense. The bigger a die is, the more chances that mistakes and errors can happen. Keeping it small means that you get more "good" chips from your foundry process.

So thus in January 2019, AMD and GlobalFoundries announced a seventh amendment to the wafer supply agreement. It allowed AMD full flexibility to purchase 7nm node and beyond without any one-time payments or royalties. It also set purchase commitments for GloFo's 12nm nodes and above - Zen 2 would still use 12nm for I/O die.

Since then, TSMC has been AMD's main foundry for its leading edge chiplet processors. When Zen 3 was released nearly two years later, an enhanced version of TSMC 7nm and GlobalFoundries 14nm again fabbed the chip's die components. TSMC also fabbed the chips for the latest Radeon RX series of graphics cards - codenamed "Navi". And TSMC's 5nm process is anticipated to fab AMD's upcoming next Zen 4 architecture, scheduled to launch some time next year.

Conclusion

While there is some initial talk about AMD potentially second-sourcing with Samsung Foundry, nothing has been announced and AMD remains one of TSMC's closest and most important partners. Apple and Huawei almost always gets the leading edge for the first year or so, but AMD follows thereafter.

AMD's path transitioning away from its own manufacturing to GlobalFoundries and then to TSMC has been a rocky one. I think it demonstrates the difficulty of the whole venture and why so many other companies have dropped out in the race. Now there are just three leading edge chip makers left, and just two of them do work for other customers.

And for what it is worth, GlobalFoundries has not really looked back ever since the 7nm halt. They seem to be doing well financially and on track for an eventual IPO. There is a big lucrative business in serving customers looking for just-good-enough.

AMD's market share has returned to its highest levels since 2007, its products praised for outperforming their Intel rivals. Intel is starting to get itself unstuck after losing years to manufacturing delays at its 10nm and 7nm level. AMD will continue to rely heavily on TSMC being able to produce enough acceptable leading edge product to continue AMD's return from the brink.

Follow Asianometry's youtube channel

TAGS: mdi-pound TSMC
About the Author
Asianometry
Author >

The Asianometry YouTube channel covers technology, economics, and history through the context of Asia. Recent videos have covered semiconductors, EV batteries, and development economics.

Follow us on social media
icon-share-facebook Facebook
icon-share-twitter Twitter
icon-share-linkin Linkedin
icon-share-newsletter Newsletter
Partners