The matter of receiving legs is not an overnight achievement, and in practice, it is also quite difficult.
”
Author | Wang Yixiang
Edit | Li Shuaifei
Intel, a world-class semiconductor giant at 54, is undergoing a painful and difficult transformation.
This transformation has actually been going on for many years, because in the past decade, Intel has had to face a core problem: in the long-term ADHERENCE to the IDM model, the development of its chip manufacturing technology and business has not kept up with the changing needs of the times, and thus there is a serious delay, affecting the company's overall development process, especially in the process process.
This kind of constraint is no less than a long-distance runner who breaks one of his legs.
In January 2021, the transformation ushered in a major turning point when Pat Gelsinger, a semiconductor industry veteran who served as Intel's CTO, was named Intel's CEO. Today, the appointment has been more than a year.
Judging from the situation of Pat Gelsinger in the past year, his work has focused on promoting the transformation of Intel's chip manufacturing business, such as announcing the IDM 2.0 strategy in March last year, establishing a relatively independent wafer manufacturing division IFS (Intel Foundation Service) and announcing multiple partners, formulating new process evolution plans, investing in new wafer manufacturing plants, and promoting cooperation between Intel and TSMC in chip foundry.
It can be said that almost all of the efforts made by Pat Gelsinger in office are trying to connect this "broken leg", the core of which is IDM 2.0.
Of course, the matter of taking over the legs is not an overnight achievement, and in practice, it is also quite difficult - this is not only a technical problem, but also a business model problem, but also a business development strategy problem. For Intel, this is a far-reaching change, and it will even determine the future direction of its entire semiconductor industry to a large extent.
For now, Pat Gelsinger's IDM 2.0 prescription for Intel has not yet been recognized in the capital market. At least in terms of Intel's stock price, he has not pulled Intel out of the slump of the stock price, and even Intel's current latest stock price ($48.95, February 1, 2022) is lower than when he first took office ($57.58, January 15, 2021).
In fact, there is also some discussion in the industry about how Intel should "take over the leg".
For example, ben Thompson, a well-known technology blogger, recently published a long in-depth blog post titled "The Intel Split" on his personal blog, affirming Intel's IDM 2.0 strategy; but at the same time, he also combined Intel's past experience of change, starting from the technology and business logic of the semiconductor industry, and believed that if Intel wants to revitalize its strength, it should split its chip design department and chip manufacturing department and let the two develop independently. While its views are refreshing, they are also quite thought-provoking.
Based on this, Leifeng Network has compiled this blog post without changing the original meaning, and the following is the content of the text:
Intel's latest earnings report won't be released until next Wednesday — but it seems to me that whatever CEO Pat Gelsinger announces to the outside world, the real news about Intel comes from TSMC.
According to the Wall Street Journal, TSMC said it will increase investment this year to increase production capacity by 47% to cope with the surge in demand caused by the global chip shortage, and the capital expenditure budget will reach 40 billion to 44 billion US dollars, a record high - last year's figure was 30 billion US dollars.
For this scale of investment, Bloomberg reporter Tim Culpan commented that this is a "warning" to Samsung and Intel, which are also fabs:
From a technical point of view, Samsung is the closest opponent to TSMC; however, Samsung also produces screens and invests a lot of semiconductor-related investments in the production of memory chips that TSMC does not involve, and the two cannot be equal.
Intel decided to join the ranks of wafer foundry last year, in addition to manufacturing chips for its own brands, it will also provide chip foundry services for third-party customers in the future, in order to compete with foundries such as Samsung and TSMC.
But Intel lags behind these two companies in terms of technical strength, and ironically, Intel's high-end chips will also be handed over to TSMC in the future. Pat Gelsinger may be confident that Intel will catch up with the pace of the leader, but it seems to be far behind in terms of capacity and capital investment.
In fact, what is worse is that intel, after becoming a customer of TSMC, while reducing the scale of its own chip manufacturing, it is also equivalent to this part of the benefits ceded to competitors, making TSMC more powerful.
1
Kissinger's chip design path
I love a quote from Michael Malone in The Intel Trinity: Moore's Law is more of a choice than a law:
Moore's Law is a social contract between the semiconductor industry and other industries, that is, the semiconductor industry continues to strive to maintain the operating rules of Moore's Law for as long as possible, while other industries pay for the results of the former's rapid development. Moore's Law doesn't work because it's inherent in semiconductor technology; rather, if one morning the world's greatest chip companies decide to stop technological progress, Moore's Law will be immediately abolished that night, and all its effects on Moore's Law will disappear within the next few decades.
Moore's Law came into being in 1965, and in the next 50 years, these choices fell into the hands of Intel, and one of the main decision makers was Pat Gelsinger.
He joined Intel after high school, worked on the 286 processor project during his undergraduate studies at Stanford University, then led the 386 processor project during his master's degree, and by the time he graduated with his master's degree, he was already the boss of the 486 processor project— he was only 25 years old at the time.
Intel 486 processor
At that time, Intel was already a vertically integrated manufacturer (IDM), from chip design to manufacturing, from chip design to manufacturing, completely different from NVIDIA, which is only responsible for IC design or TSMC, which is only responsible for foundry manufacturing - but now, as other companies focus on different parts of the chip manufacturing process, this degree of integration has been decreasing.
In the 1980s, however, Intel also had to tackle many new challenges, including chip design, which Pat Gelsinger elaborated on in a 2012 paper titled "Coping with the Complexity of Microprocessor Design at Intel—a CAD History":
In a 1965 paper proposing Moore's Law, Gordon Moore said that the rate of growth in the number and performance of chip transistors that he predicted was unlikely to be sustained because of the increasing complexity of chips and the need for product design that could keep up with his predictions. However, the fiercely competitive business environment has driven the development of technology, so that the number of available transistors for chips in the new generation process has doubled every two years compared with the previous generation, and the number of transistors has even increased by 4 times in the years when the architecture has changed. In this process, Intel's microprocessor design team needs to find a way to keep up with the rapid progress of the process.
This exponential rate of growth is not designed to be achieved by multiplying engineers—in fact, to achieve it, new design methods and innovative automated design software must be introduced into the design of each generation of processors. These methods and tools have always followed principles; for example, improving design abstraction and becoming more precise in circuit and parasitic modeling, while using increasing levels, regularities, and automated synthesis.
Often, when a task becomes too painful to perform using the old methods, people come up with a new approach and related tools to solve the problem. In this way, tools and design practices continue to evolve, always solving the most labor-intensive tasks at hand. Of course, tools evolve from the bottom up, from layout tools to circuits, logic, and architecture. Usually, at each level of abstraction, the verification problem is the most painful, so it is solved first, and the synthesis problem at that level is solved later.
This feedback loop on design and manufacturing is necessary for cutting-edge innovation — as Clayton Christensen notes in his book Innovator's Solutions:
When product functionality and reliability cannot meet the needs of customers at a specific market level, companies must compete to make their products as best as possible. In this competition, companies that build products around proprietary architectures have an advantage over those that rely on modular architectures, because the standardized nature inherent in modular architectures deprives engineers of their design freedom and prevents them from further optimizing product performance.
This is also what Kissinger did in the Intel 486 processor project:
The 386 processor largely followed the logic of the 286 processor, but at the 486 processor, the design was much more radical, moving to a pipelined design, integrating floating-point units, and adding an 8k capacity on-chip cache. Chip design is under pressure due to fewer continuations of previous-generation designs and a 4x increase in the number of transistors. We were going to solve the problem by piling people up, but the problem was that we had the ability to afford the team, find them, train them, and manage them effectively. In the end, we assembled a design team of just over 100 people for the 486 project... To execute this visionary design process, we also built a CAD system that didn't exist.
In order to build new chips, Intel at that time always had to explore a new set of methods and means as part of its entire efforts from design to manufacturing.
2
Intel's rigidity
However, 30 years have passed, and Intel has no longer maintained a leading position in chip manufacturing - instead, TSMC, which is not involved in IC design, has become a manufacturing leader.
In this regard, TSMC founder Zhang Zhongmou once told the Computer History Museum:
In fact, when I was working at Deyi and General Instruments, I saw many IC designers have the idea of leaving their own chip companies, but the biggest and only obstacle preventing them from leaving to start a business is that they can't raise enough money to set up their own company.
Because at that time, the general view in the industry was that every semiconductor company needed to make its own wafers, which was the most capital-intensive part of the semiconductor or integrated circuit industry. I noticed that these people all wanted to leave but gave up due to a lack of ability to raise a lot of money to build a fab. So I thought, maybe TSMC as a pure foundry can solve this problem. Ultimately, TSMC's presence gave those chip design engineers the opportunity to successfully form their own design companies and become our customers, who built a stable and growing market for us.
It's also as Clayton Christensen puts it in his book The Innovator's Solution:
Once customers' needs for functionality and reliability are met, they begin to redefine what is not good enough. What's not good enough is that they can't get what they really need as easily as possible. They are beginning to be willing to pay high investments for higher performance and improvements in production speed, convenience, customizability, and so on. When this happens, the competitive basis of the market changes.
Today's TSMC is willing to oem chips for companies that do not have wafer manufacturing capabilities - it is worth mentioning that today's modular chip development and manufacturing process is inseparable from Intel, especially kissinger's tool system built in the era of the 486 processor project. In the above paper, he said:
All of these tools combined to form a system called RLS, the first RTL layout system for major microprocessor development programs... RLS succeeds because it combines three basic elements... Each of these elements in isolation is not enough to revolutionize design productivity, but must be a combination of all three. These three elements were later standardized and integrated by the EDA industry. This system became the foundation of all ASIC industries and the universal interface for the fabless semiconductor industry.
The problem is that Intel, which is accustomed to making its own set of tools and processes, has gradually become disconnected from the industry.
Although Intel has also worked with EDA providers such as Synopsys and Cadence, most of its chip design process is done on Intel's own tools and has made targeted optimizations to its own foundries. This makes it difficult for other third-party companies to choose Intel to build their own chips in the past and become Intel's third-party custom foundry customers.
Worse still, Intel has wasted a lot of time building tools that used to be good for differentiation, but are now standardized.
In this regard, in March 2021, Kissinger mentioned in Intel's Chip Foundry Services (IFS) related announcement that it supports Synopsis and Cadence, which is a good thing - if even industry-standard design tools and IP libraries are not supported, Intel's chip foundry services will be difficult to gain market favor.
Last year Kissinger mentioned that IFS was just one component of Intel's IDM 2.0 strategy, which is exactly what I appreciated.
Intel manufactures and manufactures some of the chips developed by this brand (IDM 1.0);
Some high-end chips are produced by third-party foundries such as TSMC;
Then there's IFS, intel's chip foundry services for other customers.
I've been calling for a division like IFS for a long time, and even last year, a month before Kissinger's speech, suggested he split Intel's corporate architecture — and the proposal of IDM 2.0 suggests that it may not go that far, which is understandable given the historical inertia that IDM 1.0 brings.
But as I thought more and more, and considered its relevance to other sectors, I began to wonder if I was closer to the proposal I had proposed [to spin off Intel] than I realized.
3
Microsoft and Intel
In 2018, in an article titled "The End of Windows," I traced the remarkable transformations that Microsoft has taken under the leadership of Satya Nadella.
Satya Nadella's first public event as CEO of Microsoft was the release of the iPad version of Office (which was actually a project initiated by his predecessor Ballmer, which has not been released due to the delay of Windows Touch). What's really important, though, is that Satya Nadella made a big splash and used it to signal to the rest of the company that Windows is no longer a necessity for Office.
Not only that, but in the same week, he renamed Windows Azure to Microsoft Azure — sending the same message.
When Pat Gelsinger talked about choosing TSMC as an Intel foundry chip, I immediately thought of the above. In fact, the decision on TSMC's foundry was actually initiated by Bob Swan, the previous CEO of Intel. However, the outside world believes that Intel's choice of TSMC FOUNDRy is a stopgap measure, and once it makes a technological breakthrough, it is bound to compete with TSMC in terms of wafer foundry.
In fact, this is also the idea hidden at the beginning of this article.
At the same time, this is why I said at the beginning that TSMC's expansion of capital expenditure is an important piece of news – the main impetus for TSMC's expansion of capital expenditure seems to come from Intel. Reports from Digitimes show:
According to industry insiders, TSMC plans to set up a new production base in the Baoshan area of Hsinchu, northern Taiwan, to produce 3nm chips for Intel. TSMC converted part of the site to a 3nm process manufacturing facility called P8 and P9, which was originally designed for the Sub-3nm Process Technology Research and Development Center, the sources said. P8 and P9 at TSMC's Baoshan plant will each be able to process 20,000 wafers per month and will be committed to fulfilling Intel's orders, the sources said.
The source pointed out that TSMC intends to distinguish its Intel chip production from Apple's chip production, so it decided to separate its 3nm process production line dedicated to meeting the orders of these two major customers, also to protect the respective confidential products of the two major customers. Sources pointed out that Intel's huge demand may be enough to persuade TSMC to revise its manufacturing blueprint for pure wafer foundries, and the partnership between the two sides may also be long-term.
As the two giants of hardware and software, Intel and Microsoft are bound by their history, and their business layout is built on the two opposite ends of the computer in the software and hardware dimensions, the former is the atom, the latter is the bit. Both vendors have an unshakable belief in the foundation of their business model — for Microsoft, windows' influence on PCs, servers, and more; for Intel, wafer fabrication was once its temporary advantage.
However, before Nadella changed Microsoft's development strategy, he had to get rid of the legacy of Windows — and now, Kissinger wants to make the same changes at Intel, which means that he has to make changes to his own manufacturing plant.
Look at the EDA problem mentioned above: In the past, it was difficult to convince Intel's engineers to abandon their own solutions to adopt industry-standard solutions, because only potential foundry customers benefited (but not Intel itself), which was one of the main reasons for the failure of its custom foundry business.
However, if Intel wants to rely on TSMC to manufacture chips, it has no choice but to align with industry standards.
In addition, just as Windows needs to play its own advantage to win the competition rather than rely on the support of Office or Azure, the manufacture of high-end x86 chips in the future is no longer monopolized by Intel's own factories, and Intel's foundry business must face the competition with TSMC - not only to win third-party customers, but also to win orders from its own chip design team.
4
Split Intel
In the past, when I talked about how Intel should be split, I focused on the motivators in it:
The integration of chip design and manufacturing has been Intel's moat in the past few decades, but now this integration has become a shackle hindering the development of the two major business units (design and manufacturing), on the one hand, Intel's chip design is bound by manufacturing factors such as process, on the other hand, this integration is not conducive to incentivizing the manufacturing sector.
In the field of chips, design often has higher profits, such as Nvidia's gross profit margin between 60% and 65%, and the TSMC gross profit margin responsible for OEM Nvidia chips is close to 50%, due to the first-hand design and manufacturing, Intel has traditionally a profit margin closer to Nvidia, so Intel's manufacturing department always gives priority to its own chip, which means that the service to other potential customers may be worse, and it also means that there is a lack of continuous improvement to meet customer needs. It also involves the issue of customer trust: Are Intel's friends at ease handing over their chips to rivals for OEM? Especially when Intel prefers to prioritize its own chip manufacturing needs.
The only way to solve this problem is to divest Intel's manufacturing operations. Of course, building collaboration with third-party customers takes time technically, not to mention a huge IP library. However, based on the motivation to survive the competition, Intel's manufacturing division is not difficult to complete such a transformation after becoming an independent enterprise.
Obviously, Intel has not been broken up, but from the fact that TSMC has set up a factory specifically for its investment, Kissinger has recognized the shackles of the existing business architecture and made every effort to get rid of it.
So, it seems increasingly clear that the goal is to "dismember" Intel (De-Integrate)—that is, Intel, as a design company, is essentially fabless-free, handing over its business to the best foundries in the world, whether or not intel itself.
At the same time, Intel as a manufacturing company must fend for itself (of course, having exclusive access to x86 IP blocks can serve as a good bargaining chip to attract independent chip design companies), even Intel's own CPUs.
5
Kissinger's Grove moment
（Grovian Moment）
It's unclear whether Intel's strategic transformation will be successful, but given the high cost of building a fab and the importance of keeping it running at full capacity, it's clearly risky, and outsourcing a portion of its manufacturing operations will make TSMC even stronger (after all, TSMC has another big customer beyond Apple).
However, judging by Intel's performance over the past decade, maintaining the status quo may be more risky, as a troubled manufacturing sector with a process level can drag down chip designs.
This reminds me of the story of Moore and Andy Grove who transformed Intel from the memory business:
Intel started out as a company in the memory business, and became famous in the industry with the launch of the first metal-oxide semiconductor SRAM and the first commercial DRAM, and it was the memory business that brought Intel its first pot of gold, when they invested their best employees and equipment in memory, and firmly believed that memory was its technology driver, allowing all other businesses, including the fledgling microprocessor, to start. As Andy Grove puts it in his book Only Paranoia Survives, "Our priorities are defined by our cognition, after all, memories were us." ”
The problem is that in the mid-1980s, competitors from Japan struggled to produce more reliable memory at a lower cost with government funding.
Andy Grove once recalled:
Sometime in mid-1985, we were already wandering in confusion for some time, when I was discussing intel's dilemma with chairman Gordon Moore... I asked Gordon, "What do you think he would do if the board kicked us out and appointed a new CEO?" Gordon replied without hesitation: "He will make us abandon the memory business." I stared at Gordon a little sheepishly and said, "Then why don't we do it ourselves?" ”
Kissinger, Intel's first CTO, was once considered Intel's next CEO, but he chose to leave in 2009. In the decade since he left, Intel has struggled with its old architecture and model, no longer the same person who once built the 486 processor and the tools that went with it.
But now, Kissinger has returned to Intel — and if Intel wants to be a great chip design company and a great chip manufacturing company, then one thing has to be done: split the two.
via : https://stratechery.com/2022/the-intel-split/, compiled by Leifeng Network.
END