AI chip battlefield three key Chinese

Su Zifeng, Huang Jenxun, Zhang Zhongmou, this three-person game is also the epitome of the most contemporary characteristics of the AI industry.

Zhao Dongshan, reporter of "China Entrepreneur"

Editor|Li Wei

Header image source|Respondent

The computing power war set off by the big model, behind which three Chinese entrepreneurs became the protagonists.

On June 14, the global AI community was waiting for a press conference. At the press conference, AMD Chairman and CEO Su Zifeng stood in the spotlight wearing a blue stand-up collar jacket and smart short hair, and intensively released a number of new AI software and hardware products, including the new GPU (graphics processing unit) MI300X designed for large language models, taking the initiative to challenge the king of this field, NVIDIA.

The MI300X can speed up the processing speed of generative artificial intelligence, with up to 192GB of memory, which is more than the 120GB memory of the NVIDIA H100 chip, which means that it can train large language models with larger parameters than the NVIDIA H100 chip.

When Nvidia won the lion's share of the AI computing market with its GPUs, there was always a keen discussion about when AMD would launch a competitive product. Now, Su Zifeng has come, and has set off a storm.

This storm involved the world's three major chip companies: AMD took the initiative to invite Nvidia, Nvidia's stock rose sharply that night, and as the manufacturer of AMD, TSMC, which was "behind", received a large number of orders.

The three companies at the helm, Su Zifeng, Huang Jenxun and Zhang Zhongmou, are three of the most high-profile Chinese entrepreneurs in the global AI chip industry. This three-person game is also the epitome of the most contemporary characteristics of the AI industry.

NVIDIA's GPUs are regarded as the best products for training AI large models, accounting for more than 60% of the market share; AMD is considered by US investment banks to be NVIDIA's most powerful competitor; TSMC is a manufacturer of many chip design manufacturers such as NVIDIA and AMD, and its annual revenue accounts for 30% of the global semiconductor output value.

Since November 2022, with the advent of ChatGPT, it has triggered a boom in large-model entrepreneurship around the world. Behind the carnival of big models, the competition for AI computing power is pushing these three companies to the center of the technology stage.

Half a month ago, NVIDIA's market value exceeded $1 trillion, becoming the world's first chip company with a market value of more than $1 trillion; Not so long ago, Buffett was at Berkshire. At the Hathaway annual meeting, TSMC was praised, saying that no company in the chip industry can match it.

The global chip industry mainly follows the following three development models: from chip design to manufacturing, all independent completion, such as Intel and Samsung; Only do chip design and research and development, manufacturing is done by foundries, such as AMD, Qualcomm, NVIDIA, etc.; Focus on completing OEM manufacturing for chip design companies, and do not do their own design, such as TSMC, SMIC, etc.

Su Zifeng, Huang Jenxun and Zhang Zhongmou, starting from different places, jointly walked to the center of the world's AI chip battlefield.

Su Zifeng: Lead AMD back to life

Su Zifeng has the confidence to attack Huang Jenxun. In the past decade, she has personally saved AMD, which was once on the verge of bankruptcy, and increased AMD's stock price nearly 30 times in less than a decade.

This is a company that was born before Silicon Valley. In the early 80s of the 20th century, AMD began to manufacture microprocessor chips for IBM, and once surpassed Intel's autonomous processors, accounting for about a quarter of the server chip market. However, the good times did not last long, and when Su Zifeng took over in 2014, AMD was saddled with $2.2 billion in debt and even had to sell its own park and rent it instead.

In 1969, the year AMD was founded, Su Zifeng was born in Tainan City, Taiwan, and immigrated to the United States with his father at the age of three. In 1986, 17-year-old Su Zifeng was admitted to the Massachusetts Institute of Technology to major in electrical engineering. When asked why she chose this major, she said: "Because I heard that it is the most difficult major. Su Zifeng began studying silicon technology in her sophomore year, and often went to various factories to learn and apppile the chip manufacturing process during college. At MIT, Su Zifeng studied until her Ph.D.

After graduating with a Ph.D. in 1994, Su Zifeng joined Texas Instruments Semiconductor Process and Components Center as a technical specialist. A year later, he joined IBM's R&D department, responsible for developing copper chip manufacturing processes, and then served as the head of IBM's R&D department and special assistant to the CEO, and worked at IBM for 13 years.

Su Zifeng showed her talent in the field of chip design very early. While working at IBM, she helped design semiconductor chips that used copper circuits instead of traditional aluminum circuits as a researcher, making the chips run 20 percent faster, and IBM executives quickly discovered her talent. In 1999, a year after the introduction of copper circuit technology, IBM CEO Louis Gerstner decided to make her a technical assistant.

In January 2012, Su Zifeng joined AMD and has served as Chief Operating Officer, Senior Vice President and General Manager of Global Business. In June 2014, Su Zifeng became CEO in the restructuring of AMD, in charge of AMD's large marketing business, and also became the first female CEO in AMD's history.

Su Zifeng is affectionately known as "Su Ma" by AMD engineers. On the second day of becoming CEO, Su Zifeng gave a speech to frustrated AMD employees on an all-staff conference call. "I believe we can build the best product possible," she says. ”

This is the first step in Su's overhaul of AMD, and she already has three things in mind: create great products, deepen customer trust and simplify the company. "Just do those three things to keep it simple," she said, "because if it's 5 or 10 things, it's hard." ”

AMD's share of the server market continued to decline, but Su Zifeng decided to focus most of his energy on designing chip products. She decided to prioritize the development of a new chip architecture called Zen.

This decision was a success when Zen finally launched in 2017. By the time the third-generation Zen was released in 2020, it had become the market leader in terms of speed. The Zen architecture is now the foundation for all AMD processors. Even when AMD had no chips to sell, Su spent years explaining to customers what he was doing.

Opportunity tends to favor those who are prepared. In 2019, Intel was in trouble due to production delays and the loss of Apple, a customer, Su Zifeng keenly seized this opportunity and led AMD to win customers such as Lenovo, Sony, Google and Amazon, and also allowed AMD to return to the top.

In 2021, Su became the first woman to win the Robert N. Noyce Award, IEEE's highest semiconductor award. In 2022, she led AMD's $48.8 billion acquisition of chip company Xilinx, which makes programmable processors that help speed up tasks such as video compression.

As part of the deal, Xilinx CEO Bloomingbale became AMD president and head of AI strategy. This also allowed Su Zifeng to challenge Huang Jenxun and have more chips.

In June 2020, an Associated Press survey showed that Su Zifeng was paid US$58.5 million (about RMB 419 million) a year, including US$1 million in basic salary, US$1.2 million in performance bonuses, and US$56 million in stocks. Su Zifeng became the first woman among the world's most profitable CEOs and an indispensable key figure in the chip market.

Jensen Huang: Enter the trillion dollar club

The year Su went to the United States, 9-year-old Huang was also sent by his family from Taipei to the United States to attend a rural boarding school in Kentucky.

This is a dangerous lesson in the life of young Wong Jen-hun. He later recalled: "The school was actually more like a juvenile correctional home, where every child had a knife and classmates were tattooed all over their bodies. Instead, however, Wong learned to be strong and adaptable.

Two years later, Huang entered a regular school. In addition to his academic performance, he also won the third place in the national doubles in table tennis. But he loved technology and entered the University of Oregon at the age of 16 to study electrical engineering, and aspired to become a global graphics emperor during his college years.

After graduating with a bachelor's degree, Jensen Huang received his master's degree from Stanford University. Later, he went on to work as an engineer at AMD and LSI Logic. In 1993, at the age of 30, Jensen Huang and two engineers co-founded NVIDIA.

Different from the chips made by Intel and AMD, Huang Jenxun paid more attention to the graphics chip that allowed games and images to run smoothly, because at that time, personal computers had just entered the home, but entertainment functions were still missing, unable to meet the needs of game running. But at that time, no one in the outside world was optimistic about the new direction of GPU.

Before 1999, Huang had introduced two chips. But because of betting on the wrong technology, Huang exhausted the company's early investment, shrinking from more than 100 to more than 30. At the end of 2000, NVIDIA acquired its old rival 3DFX for $110 million. That same year, Huang challenged chip giant Intel.

Benchmarking Intel founder Gordon Moore's "Moore's Law", Huang Jenxun proposed "Huang's Law", that is, NVIDIA's products are upgraded every 6 months and the function is doubled, which is 2 times faster than "Moore's Law".

In late 2006, Huang opened up GPUs to software developers and launched the CUDA platform, which allows developers to use the computing power provided by NVIDIA for purposes other than graphics. Although it had little effect at first, with the advent of the era of artificial intelligence, developers soon realized that GPUs were excellent at supporting the complex calculations of modern AI systems.

Later, NVIDIA GPU and CUDA programming language became the basis and standard for AI development and training. This surprised Huang. In a 2016 interview with Forbes, he mentioned that he had expected GPUs to be used outside of gaming, but never thought of turning to deep learning applications.

Source: Visual China

Today, NVIDIA's products have become the hard currency in the global AI computing power. Huang personally handed over the world's first AI supercomputer, DGX, to OpenAI, which developed ChatGPT. Since then, the competition of computing power of global AI companies has also become a competition of how many NVIDIA GPUs.

In the past three years, NVIDIA's market value has soared from $150 billion to $1 trillion, and Huang Jenxun is also known as the "godfather of AI". According to the estimation of market research agency Jibang Consulting, 60%~70% of the more than 1.2 million AI servers shipped this year are equipped with NVIDIA's GPUs.

Zhang Zhongmou: Produce 30% of the world's computing power

In 1997, while Huang was still struggling to find a chance to start a business, another entrepreneur from Taiwan, China, reached out to him.

At that time, Huang had just developed the third-generation product RIVA128 and was approved by Microsoft. But the next challenge before him was how to produce mass production quickly. At the time, it cost $100 million to build a fab, which is astronomical for a startup.

TSMC founder Zhang Zhongmou promised to help Huang Jenxun produce the original production.

When Huang was 1 year old, Zhang Zhongmou had already received a doctorate from Stanford University's Department of Electrical Engineering. Interestingly, after graduation, he intended to work for Ford Motor, because Ford paid him a monthly salary of $1 less than a chip company, he chose a chip company, and then entered Texas Instruments, and was promoted to general manager of the integrated circuit division the following year, and then worked here for more than 20 years. In 1985, after losing the election as CEO of Texas Instruments, 54-year-old Zhang Zhongmou returned to Taiwan from the United States and founded TSMC.

While still working at Texas Instruments, Zhang had a radical idea: When chip demand rises, chip design and manufacturing should be separated, because the companies that design chips lack the expertise to produce semiconductors. As technology advances and transistors shrink, manufacturing equipment and R&D costs will rise, and only companies that produce a large number of chips have a cost advantage.

However, at that time, Texas Instruments, Intel, and Motorola were all self-developed and self-produced. AMD founder Jerry Sanders even shouted the classic quote, "A real man has a fab." Zhang's ideas have not received any support, and more importantly, some chip companies will worry that their design ideas and ideas will be copied.

But in the end, it turned out that real men can indeed live without a fab.

Source: Visual China

In order to dispel the doubts of chip design companies, Zhang Zhongmou promised, "TSMC will never design chips, only make chips." Coupled with the chip design revolution at the time, which made it easy to separate design and manufacturing, Zhang Zhongmou caught up with the perfect opportunity, and the industry even compared it to the chip industry's "Gutenberg moment".

Today, the new wave of chip entrepreneurs in Silicon Valley doesn't have a fab.

Chips from NVIDIA, AMD, Apple and other companies are mainly manufactured by TSMC. It is precisely because of the choice of many chip design companies such as NVIDIA that TSMC's brilliance has been achieved. In 2017, TSMC's market value surpassed Intel and became the world's largest chip manufacturing company. Also in the same year, Zhang Zhongmou announced that he would retire the following year and would no longer hold any position at TSMC.

According to the "TSMC Business Report to Shareholders", TSMC's revenue accounted for 30% of the global semiconductor (excluding memory chips) output value in 2022. In terms of shipments, in 2022, TSMC provided 288 different process technologies and produced 12,698 different products for 532 customers.

After Zhang Zhongmou retired, TSMC rode the AI east wind to leave the competition far behind. His successors are increasingly aware that as the industry begins to enter the AI era, a smarter and more connected world will create a strong demand for computing power and low-energy computing, and will welcome them with a larger market.

Resources:

"Chip Wars", Chris Miller

"Su Zifeng saved AMD, and now she wants Nvidia's AI crown" Forbes

"Huang Jenxun: The person who single-handedly created the "new king" of the chip market", Interface News