We were “trapped in place” for nearly a year. However, one year ago we might not have thought that the “transformation thinking” that this state of affairs brought to large and small companies in China and even the world would evolve from a handful of flames in January to one that burned all over. The digital fire of the industry. “During the epidemic, the technology industry has actually decoupled from the overall economy. The semiconductor industry in particular is still booming.”
Liu Deyin, the current chairman of TSMC, at the June 2020 shareholder meeting proved that this fire consumes the “underlying fuel”—the speed and appetite of chips, which have also become unprecedented.
Three months later, TSMC raised its 2020 revenue forecast for the second time. The main judgment of the world’s largest semiconductor manufacturer comes from the strong demand for 5G mobile devices and high-performance computing equipment in data centers.
But on the other hand, some semiconductor giants serving traditional manufacturing such as automobiles and industries are under tremendous pressure to rapidly shrink revenue.
Therefore, at this very special time node, there has never been a “relentless” semiconductor market without the word “sympathy”. In 2020, it will usher in its destined era of great integration.
In 2020, in less than 5 months, the semiconductor industry has produced 5 mergers and acquisitions with strong alliances (pictured below). Every semiconductor company involved is well-known, and each amount exceeds tens of billions of dollars. Each merged company will become a new semiconductor behemoth.
As of November this year, the world has reached at least US$115 billion (760.3 billion yuan) of acquisitions in 2020. Whether it’s the single maximum amount or the total amount,Have broken the semiconductor industry M&A records maintained in the past 20 years。
Not sure about the current situation, keep warm
To a certain extent, the frequent M&A wave of huge transactions in 2020 is more like a continuation from 2015 to 2016.
According to IC Insight’s monitoring data, the global semiconductor M&A amount in 2015 reached 103.3 billion U.S. dollars, and this trend continued in 2016. The total transaction value exceeded 98.5 billion U.S. dollars, almost five times that of 2010-2014 (average annual The transaction volume is only 18.6 billion US dollars).
Industry insiders analyzed at the time that as the global macroeconomic growth slowed, the growth of the semiconductor industry also slowed down, but R&D and capital intensiveness continued to increase, competition became increasingly fierce, and mergers and acquisitions in the semiconductor industry would be the general trend.
Li Xiaomin, chairman of Sembcorp Nano, once told Huxi that the overseas semiconductor market is highly mature compared to the domestic market, especially the analog chip market, and even the outcome has been divided. Therefore, each segment has gradually entered the stage of merger integration and oligopoly.
In 2020, the second in the analog chip market, Arnold, merged with the seventh largest Maxim, and SK Hynix merged with Intel’s storage business to become a flash memory giant second only to Samsung.
“According to previous statistics, Nasdaq could find the names of a hundred semiconductor companies ten years ago, but now there are only about forty companies left. However, while the number of listed companies has decreased, the overall revenue and market value of the industry have declined. It’s greatly improved again.”
On the other hand, making up for shortcomings through mergers and acquisitions and diversifying business to diversify business risks are more obvious in the two transactions between NVIDIA and AMD.
Advantages Nvidia, which is only fixed on AI accelerators for a long time, can obtain chip design authorizations for more than 80% of the world’s smart phones and thousands of home appliances through ARM.
Xilinx’s financial data (pictured below) tells us that orders from the aerospace, defense, industrial, and surveying and mapping sectors accounted for nearly half of its revenue, and the wired and wireless equipment sector also accounted for 30% of its sales. And AMD’s share of these markets is very low, or even nonexistent.
However, there are more new variables in 2020.
First of all, the US uses the semiconductor supply chain as a weapon to try to disrupt this highly globalized division of labor, which also adds uncertainty to the future trend of the semiconductor industry.
Although the overall strength of China’s semiconductors is not strong, lagging far behind the United States, South Korea and Japan, China is the world’s largest semiconductor sales market, with more than 50% of global sales of chips.
Therefore, the “Huawei ban” in 2019 has caused US chip suppliers including Maiwell, Inphi, Micron and other Huawei to be hit successively, and they have lowered their quarterly earnings and sales expectations.
Secondly, it is the status of a major semiconductor consumer that gives China the power to say “no” to all large-scale semiconductor mergers and acquisitions in the world.
In 2018, Qualcomm abandoned the acquisition of NXP due to China’s “visit refusal”. In other words, China’s “nod” will be an important part of whether these five huge mergers and acquisitions can reach the final step in 2020.
The game between the two countries in the semiconductor market pulls all the companies in this chain, and none of them can stay aside. Therefore, in order to get through this extremely unstable rainstorm night, but I don’t know when it will end, it is really the best policy to “keep a group for warmth”.
All want to be the “next Intel”
If you only see the above shallow reasons, then you will miss the most exciting chapter in the next 10 years of the semiconductor industry script.
One month ago, when the news of Nvidia’s acquisition of ARM was first exposed, an Nvidia engineer conveyed 4 words to Huxi to summarize the technical reasons for the acquisition:
In the development of the semiconductor industry for more than half a century, every time we want processors to have more computing power, there is basically one answer: just add more “silicon”.
However, the complexity of calculations, the limitations of the transistor arrangement on the chip board, and the cost have already begun to erode this criterion, and this is also one of the key reasons for the demise of Moore’s Law.
A senior expert who had worked in ASM and semiconductor materials companies told Huxi that in 2010, Intel reduced the chip line width to 20 nanometers, which was the limit that the lithography equipment could withstand at that time. According to its disclosure, Intel tried lithography machines from many top companies, including Asmarck and Nikon, but it still couldn’t solve the problem.
“Now, the 14 nanometers and 7 nanometers you hear are no longer the size calculated strictly according to Moore’s Law. In order to continue Moore’s Law, including Intel, all companies must make architectural innovations in chips.”
In other words, the golden law that determines the innovation cycle of semiconductor products-Moore’s Law has long existed in name. The absolute size of chips is shrinking at a slower rate and is approaching the “limit of chip manufacturing equipment”; while the looming potential of IOT (Internet of Things), artificial intelligence, and autonomous driving applications are in urgent need of new ammunition for the right medicine.
Intel has already realized this, so it challenged its own chip architecture many years ago.
Jim Keller, a chip design genius who has left, and former head of Intel’s silicon engineering department and senior vice president, participated in the design of Intel’s 3D stacked chip product Lakefield. He highly praised the use of vertical construction to redesign chips.
Simply put, this method can stack chips with different functions together, and transmit data vertically upwards through the bottommost block, allowing high-speed interconnection between chips.
Picture from Hot chips
Of course, Intel’s strongest rival in the PC market, AMD, is also trying the same thing-they are trying to “mix and match” a set of different functions, different processes, and even different brands of small chips (Chiplets) to combine them into one system.
For example, only the core of AMD’s second-generation enterprise-class “Xiaolong” processor uses TSMC’s advanced 7-nanometer process, while other parts use low-cost old processes, such as 14-nanometer or 20-nanometer. Because of some functions, the old craftsmanship is adequate.
In other words, the materials and packaging in the entire life cycle of a chip (material production, design, manufacturing, packaging and testing) will play an increasingly critical role in changing the chip structure.Therefore, Xilinx’s well-known 2.5D packaging technology will definitely help AMD design more complex SoCs.
Xilinx launched its data center product Versal ACAP in October 2018, a fully software programmable heterogeneous computing platform
So from a business perspective, a more realistic question is, what are the profitable opportunities that drive the giants to scramble for chip architecture innovation?
We might as well answer this question first-why do Google, Microsoft, Amazon, and Ali continue to participate in and design chips themselves, and use the most advanced manufacturing and packaging processes?
The answers to both questions point to one market: data centers.
Although we do not comment on the feasibility of artificial intelligence commercialization, in just three years, the application penetration rate of artificial intelligence in various industries has far exceeded the public’s imagination——
As small as Toutiao’s and Taobao’s personalized recommendations, as large as Tesla’s assisted driving function and the predictive maintenance of factory production lines…In the digital transformation process of all industries such as finance (banking), retail and industry, there are Artificial intelligence algorithms are hidden deep in it.
Therefore, from the perspective of the overall application volume, the purchase of semiconductors driven by artificial intelligence algorithms in the next five years will not be underestimated. At a time when the magnitude and complexity of algorithm models are exploding, chip architecture innovation becomes imperative.
At the same time, with the wave of cloudification of personal entertainment and office, consumer hardware and enterprises, cloud computing giants operating super data centers and large and small enterprises with data centers will become the new high-performance processors. Buyers.
Then a lot of things are to be expected-Intel, the biggest winner with 70% market share in the data center, will definitely encounter a more fierce siege.
Although the PC market has let the public know about Intel, the “data center” is its revenue source second only to the PC market, accounting for more than one-third of total revenue. In Q4 of 2019 alone, this part of revenue was as high as 7.2 billion US dollars, while AMD’s 2019 annual revenue was only 6.7 billion US dollars.
Regarding the five huge mergers and acquisitions in 2020 (the first picture), a semiconductor expert told Tiger Sniff that from the perspective of business competition, Nvidia, AMD, and Maiwell are all targeting the data center market.
Maiwell, who just announced the acquisition of Inphi, made it clear from the beginning: “Inphi’s technology is the core of the cloud data center network. We hope to use Inphi’s unique silicon photonic materials and DSP (digital signal processing) technology to expand data. Central infrastructure market.”
In addition, Nvidia, which has become the largest supplier of AI accelerators in the data center segment, first annexed Mellanox, which can provide efficient connections for thousands of servers, in April 2020. Later, CEO Huang Renxun publicly challenged Intel when he announced the acquisition of ARM:
“I am very excited to concentrate a lot of resources to turn Arm into a world-class data center CPU supplier.”
However, ARM, which is popular in the mobile device market due to its concise and refined chip architecture, has actually tried its best to impact the data center market, but it has not been very present.
An engineer wrote in the EE journal that the reason why Intel’s Xeon processors and other enterprise-level products have been indestructible in the data center is due toThree thingsTo defend:
Marketing defense. In the 2B market, the procurement department of a company is not willing to risk being fired to choose a new product, but is more inclined to choose “the system that 90% of data centers are using.”
The ecology established by the X86 instruction set architecture. It has been the industry standard for decades and the default target of most software compilers. To allow enterprises to use new CPUs, you need to let engineers out of a world where executable files are all X86.
Hard quality. Yes, Intel used to be the leader of the most advanced chip manufacturing process, so server processors are undoubtedly the most advanced and performant.
But now, two of the high walls are being shaken and crossed by the insurgents.
First of all, in June of this year, Apple was revealed that “it will continue to use ARM processors to replace Intel’s CPUs in its Mac computers” caused an uproar, followed by the bad news of various Intel personal computer chips. .
At that time, people ignored a deeper change-this may mark the end of Intel’s 40-year dominance in the X86 era.
And if the X86 monopoly is broken, it is possible to break its monopoly in the data center.
Linus Torvalds, the father of Linux, once explained the important reason why ARM servers have not opened up the market-developers want to run the same code on the cloud as on their laptops. Because the code is written on an X86 notebook, not on a mobile phone.
Therefore, ARM notebook computers are considered likely to be a catalyst for major changes in the data center market. Because even though Apple’s share of the personal computer market is only 8%, the proportion of global software developers that this company has attracted is nearly 30%.
And another barrier to “mastering advanced manufacturing technology” was also broken in September this year with the news of “Intel 7nm chip postponed” (for specific information, please refer to the article I wrote earlier, “Intel abdicates, TSMC is king”) .This means that AMD using TSMC technology can also launch a CPU with performance that does not lose to Intel.
Of course, Intel is not trying to strengthen its advantages.
They have acquired more than three artificial intelligence chip companies. In 2015, they acquired Xilinx’s main competitor, Altera, the second in the FPGA market, for $16.7 billion, and achieved success in the data center. To a certain extent, this is also one of the important reasons why AMD decided to acquire FPGA boss Xilinx.
At the end of 2019, Intel not only launched the oneAPI plan to promote the development of heterogeneous computing in data centers, but also developed the 3D chip stacking technology and AI accelerator technology just mentioned for server-level processors.
In addition, it is interesting that although Intel resolutely discarded its Nand flash memory business, it quietly retained Optane, an advanced storage technology for data centers.
Data center, picture from medium
But it is clear that the subtle and dynamic changes that occur every day in the data center from the underlying chip, data volume to the server structure have finally opened the door to power challengers that have been closed for decades.
Whether it is Super7 (Amazon, Microsoft, Google, Facebook, Alibaba, Tencent, Baidu), which has the same strong technology and financial strength, and dozens of new processor startups that have been built in the past two years, who can have and control The software and hardware capabilities of running applications on heterogeneous computing architectures break the shackles of old technology, and whoever has the opportunity to subvert.
However, a semiconductor engineer does not think this is a winner-takes-all market that will still be monopolized by a certain company in the future.
“Because the tasks of operating on the data center are very complex and the amount of data is also very large. There are GPUs that are good at, CPUs are good at, and FPGAs (programmable processors) are good at, so there are many driving factors for expanding the data center. , Not only need a chip architecture, but a heterogeneous system composed of various types of CPU, GPU, FPGA and AI accelerator.”
In short, the future winners of the data center include both vertical chip manufacturers and heterogeneous designers who can support multiple tasks—not only can the memory for storing data be combined with CPU, GPU, FPGA and AI chips as much as possible Close to the ground, they can be “spliced” together with high bandwidth, allowing data to flow efficiently between different processing elements.
“Going to mention the indentation of 7 nanometers, 5 nanometers or 3 nanometers. For the semiconductor industry, it has become gradually meaningless.” He believes that the key to semiconductor technology has begun to be chip interconnection, storage and computing integration, and packaging. Technology transfer, any company that masters new technology can become a sharer of the cake.
“The amount of data is expanding, and the number of digitally transformed companies is still increasing, which means that the scale of the data center market is still expanding and there is no upper limit. Intel may still be the leader, but this does not mean that challengers do not have their own place. .”
Chinese companies need to be cautious and calm
In the new round of mergers and acquisitions in 2020, some Chinese semiconductor players are watching movies across the bank, some are pessimistic, and some see new opportunities brought about by the technological wave.
“Whether they join forces, from a business perspective, is not really beneficial to us.” A semiconductor practitioner believes that “Bao Tuan” must have greater pricing power for customers, and it has greater say in applying for export licenses from the US government. For Chinese users, the choice is definitely narrower, and bargaining power will be further weakened.
“Of course, after the giant license is approved, some domestic companies may get the chips they want in the short term. This may be the only benefit.”
There are also semiconductor people who have similar complaints: “In the past, industrial-grade VR (a power management chip) and VRM chips could be purchased from three companies. If Arnold merges with Maxim, there may only be two now.”
In addition, Academician Ni Guangnan also pointed out in a speech in September that ARM’s chip architecture is monopolistic in some industries. If it is controlled by a US company, it will be detrimental to China. “I believe our Ministry of Commerce may reject this merger,” he emphasized.
But from a technical point of view, we should not let go of equal opportunities for innovation.
After NVIDIA announced the acquisition of ARM and AMD announced that it would acquire Xilinx, many Chinese enterprise-level chip engineers were very excited. “Look, heterogeneity has been talking for more than two years, and now it is becoming more and more certain. design!”
“You can see from Xilinx’s quarterly revenue outlook for the past year. They have begun to increase investment in artificial intelligence and other directions. At the same time, data center revenue is gradually increasing.”
A data center chip architect told Huxi that Xilinx’s FPGA chip is an indispensable product for their product verification, because the architecture adjustment is flexible. Although such chips are more expensive than CPUs and GPUs, they are critical in communication equipment. Companies such as Huawei will purchase Xilinx chips in large quantities.
“Many chip companies and cloud computing giants in China are also making continuous large-scale investments in the heterogeneous computing design of data centers. Although the basic skills gap is large, everyone has seen this trend.”
But in fact, everyone knows why Chinese companies should not miss this round of new technology trains and engage in independent research and development.
“We can’t go back anymore.” A person in the semiconductor industry lamented that the choking war between China, the United States, Japan and South Korea, the monopoly of giants, the global wave of digital transformation…all reminded over and over that China’s semiconductor industry is only left The path of “self-reliance”:
“It’s like a person suddenly gave you a knife. Even if the injury is healed, the scar will always exist, and it will faintly pain every rainy night in the future, reminding you not to forget everything that China has experienced in the two years since 2018.”
Looking at it this way, according to the eight-year “rebirth” of the rise and fall of the semiconductor industry, 2020 is just the beginning of the peak moment of the semiconductor market.
In other words, the vast majority of semiconductor companies in the world will participate in a series of jungle wars around data centers in the next 5 years, and will create products or be products in an environment where technology does not follow Moore’s law to change wildly. defeat.
There is no doubt that there are no weak people in this field.