One day recently, in an open conference room at Intel's Jones Farm in Hillsboro, Washington County, Oregon, Pat Galsinger took out a chip with blue, orange, and purple silicon gold reflective light shining on its surface.
"This is the fifth generation Xeon processor (CPU)," the Intel CEO told media reporters present. "This is the company's latest server chip, and it is also the most important piece of the chip company's AI strategy puzzle.".
CPU has always been the brain behind personal computers (PCs) and servers supporting the internet, but as Nvidia seizes the opportunity of the artificial intelligence era, GPU (graphics processor) chips for AI accelerated computing are gradually expanding market share and seizing Intel's space. In addition, Intel is facing difficulties in chip process development, which has also brought opportunities for its former follower AMD.
The decline in market share can still be regained, but the decline in technological leadership is a dangerous signal. In the past few years, server vendors and cloud computing companies have been able to show respect and understanding for Intel's slow innovation efforts. However, over time, as other chip companies accelerate their exploration of more diverse innovative products, customer patience is gradually being consumed.
Since its establishment in 1968, Intel has gone through more than half a century of ups and downs, and its current state is clearly not good. The overall market is under pressure, and its performance has been continuously declining. The Economist once commented on Intel's current predicament, "It has taken a wrong path in technology and management."
But Intel, which has always been arrogant, does not admit failure. This chip company urgently needs to regroup, but adjusting its pace is so difficult. "The semiconductor industry has always been a fast-paced and ruthless industry," TSMC founder Zhang Zhongmou once wrote in his autobiography, "Once we fall behind, it will be difficult to catch up."
But history has given this company the resilience it deserves, and the clues to counterattack are hidden in the gray white office buildings of four Intel campuses in Hillsborough, Oregon, where Kissinger is located. For the outside world, most of what Intel is doing here is still a mystery.
Oregon is located in the northwest of the United States, where the Columbia River flows and is tightly surrounded by the Pacific Coast Mountains and Rocky Mountains. It borders California and Washington to the north and south, and is the terminus of the Oregon Trail during the pioneering era of the western United States. This area is rich in forestry resources, with short summers, warm, dry, and sunny weather, while winter is an ideal location for outdoor sports enthusiasts.
Hillsborough is less than an hour's drive from the capital of Oregon, Portland, with convenient transportation. Due to its advanced education, the high-tech industrial parks scattered around Hillsborough have a similar level of prosperity to Silicon Valley and are known as the "Silicon Forest". "Compared to Seattle, which is unpredictable, I prefer it here," an Intel executive who used to work in Microsoft's cloud computing department told Interface News.
Intel is the largest private employer in Oregon, with over 22000 employees in Hillsborough, Washington County, accounting for approximately 18% of Intel's global workforce. It has a chip manufacturing, research, and operations department, and according to data disclosed by Intel in 2019, it has a total annual impact on over 100000 jobs, over $10 billion in labor income, and $19 billion in GDP in Oregon.
Under the influence of Intel, semiconductors are Oregon's largest export product - the local media "Oregon Public Broadcasting" once described it as, "If you have a dollar in your pocket, then 25 to 29 cents are due to Intel's presence in Washington County."
At Intel, Moore's Law (which doubles the integration and performance of transistors every 18 months) is its fundamental law. This semiconductor company, founded in 1968, once produced the world's first microprocessor. Under the leadership of co-founder Andy Grove, it propelled chip technology by leaps and bounds, and Intel became a well-known name. For a long time, Intel represented the top technological capabilities of the semiconductor industry.
Behind this success lies Intel's investment of over $52 billion in Hillsborough over the past 50 years, where almost every Xeon processor was born and became the "heart" of Intel's research and development.
Now, four Intel campuses stand on this land, and this is just part of Intel's massive chip manufacturing catch-up plan: two years ago, Intel proposed the "four years and five nodes" plan, which means to complete the five process nodes of 7 nanometers, 4 nanometers, 3 nanometers, 20A (1 nanometers equals 10 nanometers), and 18A within four years, returning to the leading position in the chip process.
Advanced semiconductor processes require too much investment, which is a game that cowards cannot afford. In September of this year, in a conference room near Intel's headquarters, Kissinger told Interface News that half of this plan has been achieved.
In 2022, Intel completed the $3 billion third phase of its D1X chip factory in Hillsborough. According to Intel's plan, Oregon will become one of the world's most advanced chip production bases.
The focus of advancing this plan is on Oregon. Just a year ago, Intel completed the $3 billion third phase of the D1X chip factory in Hillsborough. According to Intel's plan, Oregon will become one of the world's most advanced chip production bases, and EUV (Extreme Ultraviolet) and High-NA EUV (High Numerical Aperture Extreme Ultraviolet) lithography machines will be gradually introduced here.
The basic process of chip manufacturing is to carve fine patterns on a polished 12 inch silicon wafer, known as photolithography. Subsequently, they go through a total of over 2000 processes, including deposition, photolithography, etching, cleaning, and inspection, followed by another round of repetition, and so on. After hundreds of processes, each wafer is cut into approximately 80 stamp sized bare pieces (Die).
The lithography machine is provided by the Dutch company ASML, which is one of the greatest engineering miracles in human history, consisting of over 100000 components. Generally, the photolithography process accounts for one-third of the total manufacturing cost of chips, and takes up about 40% -60% of the entire chip production time. Moreover, the control of the environment and accuracy is much higher than other processes.
ASML is located in Hillsborough's office and is assisting one of the largest chip manufacturers in the United States in adopting EUV for mass production.
If you come out of Jones Farm Park and pass through NE Evergreen 7451 in Hillsborough, you may accidentally come across the dark blue four letter ASML logo. On the latter's official website, the lithography giant introduces, "Hillsborough is one of the largest chip manufacturers in the United States that uses EUV for mass production."
"Building and connecting a wafer fab is not just about purchasing equipment and putting them together," ASML China manager Shen Bo told Interface News. It is an industry tradition for semiconductor equipment manufacturers to open offices near the wafer fab, and ASML also needs to support customers in all aspects of chip production, including the start-up of the chip production line, process debugging after the connection, and ultimately the production of the chip.
If the manufacturing process runs smoothly, the product produced after the entire processing step is the fifth generation "Zhiqiang" chip.
Thanks to the booming development of cloud computing, the demand for high-end processors used in data center servers has been increasing, and most of these data centers use Intel's x86 architecture processors. In 2022, these products accounted for one-third of Intel's total revenue of $63 billion and the majority of its net profit of $8 billion.
Compared to individual users, server and cloud computing clients crave higher performance, but at the same time do not want cost increases due to excessive power consumption. This feature needs to be reflected in the new generation of Xeon processors.
Generally speaking, the higher the clock frequency of a processor, the stronger its performance. The so-called clock frequency refers to the stable electrical pulse generated by the number of times the processor can open and close circuit switches per second, measured in hertz (Hz), which determines how many instructions the processor can execute per second.
If the clock frequency is set too high, it may not function properly due to the transistor overheating. The solution is to increase the processor core, which allows the processor to process multiple instruction sequences simultaneously, achieving enhanced computing power while keeping the clock frequency unchanged. The highest specification model of Intel's sixth generation Xeon processor this time has 64 processor cores, which is 4 more than the previous version.
The core architecture of the fifth generation Xeon processor.
Compared to the previous generation, the fifth generation Xeon has two larger bare chips with higher density that can accommodate more processor cores, while the previous generation had four. Intel senior technology expert Allen Chu told Interface News that larger bare chips can be configured with higher cache capacity, more cores, and are more suitable for large-scale virtualization use cases, simplifying chip design. Moreover, they can also bring lower data processing latency and energy consumption.
But the process from chip design to manufacturing is far from smooth sailing. The difficulty Intel has faced in advancing semiconductor technology over the past four years is that it has always hoped to catch up with the difficult Moore's Law and meet customer needs.
In fact, the birth of the fourth generation Xeon processor, which was released in January this year, has been quite bumpy - it has been delayed at least twice since its first announcement in 2019. When designing the chip, Intel adopted more complex designs and added a lot of new features, resulting in multiple validation tests to fix defects after chip casting, and had to postpone the release date multiple times. When the fourth generation Xeon was launched earlier this year, the industry's general reaction was "long waited". During that period, Intel's data center division experienced a severe test.
Therefore, the market has seen an unusual scene: in less than a year, Intel has launched two generations of Xeon processors in a row, and customers are somewhat overwhelmed.
From the perspective of competition in the IT industry, being slow is considered unacceptable.
"We only released the fourth generation Xeon in January this year, the fifth generation Xeon in December, and there may be a next generation next year. Our speed is getting faster and faster," said Chen Baoli, Vice President of Intel Data Center and Artificial Intelligence Group. "Customers and the market are demanding faster computing power updates, and we are actively responding to this."
In addition to the chip itself, Intel also had to find a comprehensive solution to face a strong competitor.
"On the data center end, Intel is actually the most mature company in terms of ecosystem development. Although we were initially slow in AI and GPU, there should be no other company that has done more in terms of collaboration with the entire ecosystem and open source community over the years." Chen Baoli said, "We all know that a friend's ecosystem has been developed for many years, but it is relatively closed."
This argument naturally refers to Nvidia, and Intel's reason for convincing customers to use open solutions is that Nvidia's software is proprietary, which means software developers cannot freely adjust it. Intel offers open source alternatives, which some customers may find attractive.
But in the current AI market, "closed" has overcome "open". Nvidia has been building a software platform called CUDA around GPUs since 2006, investing billions of dollars to help developers more easily utilize GPU capabilities to complete various tasks, and firmly bind developers to GPUs. Now, whether OpenAI, Meta, Microsoft, Tencent or ByteDance, their AI ecology is based on CUDA.
The explosion of generative AI has driven the growth of customers in processing AI computing chips. According to Morgan Stanley, AI processing chips include GPUs from Nvidia and AMD, as well as AI computing specific chips. The overall annual sales are expected to reach around 43 billion US dollars, accounting for approximately 8% of the total sales in the chip industry. Market research firm Counterpoint predicts that due to this impact, the size of the data center CPU market is expected to shrink by 15% -18% in 2023.
Sandra Rivera, General Manager of the Data Center and Artificial Intelligence Business Unit, said she is not worried that GPUs will dominate the CPU data center market. Now, Zhiqiang has built-in AI accelerators that can help customers better handle AI computing problems.
Now NVIDIA has the upper hand, and Intel will also find ways to counterattack.
For example, corresponding to CUDA, Intel has launched a software solution called OpenVINO, which Intel claims can accelerate AI computing on "any hardware, any model, anywhere.". In recent years, Intel has vigorously promoted the "One API" strategy in heterogeneous computing such as CPUs, GPUs, and FPGAs, hoping to integrate its complex product line with a packaged platform solution and reduce customer sensitivity to differences in underlying hardware.
In addition, the built-in AI accelerator can also make Intel processors more attractive in the AI field than AMD. "Unlike AMD, Intel Xeon is a CPU with built-in AI accelerators, but AMD does not have built-in accelerators. They cannot participate in software work and cannot perform data service applications," said Lisa Spelman, Vice President and General Manager of Xeon's Product and Solutions Division at Intel.
In the coming years, AI chips are expected to drive the continuous growth of the semiconductor industry. According to the prediction of International Business Strategies (IBS), a chip consulting company, the chip industry revenue is expected to double to about $1.1 trillion by the end of the 1920s, driven by advances in 5G networks, autonomous vehicle and other technologies. Now, IBS has raised its 2030 revenue forecast by $150 billion to around $1.25 trillion.
However, Intel is facing a more diverse and complex market environment than in the past, with more competitors entering the market. In addition to Nvidia, large cloud computing companies such as Amazon and Microsoft also invest in AI chips. With their massive scale, these large technology companies have the ability to design their own AI chips for their data centers and outsource the production of these chips to companies such as TSMC. These cloud computing companies are using self-developed chips to better cope with the competitive needs of the cloud service market, and also to break away from their dependence on NVIDIA. This trend is mixed for Intel.
At present, there is still room for growth in the recognition of Kissinger's changes in the capital market, and it will take time for Intel to regroup. In the past year's semiconductor market AI market, Intel's stock price has only risen by 87.99%. During the same period, Nvidia rose more than 2.5 times, and the Philadelphia Semiconductor Index also recorded a 70.7% increase. Facing the huge opportunities in the semiconductor industry in the AI era, despite the numerous challenges, this is still an unbeatable battle for Intel.