The server chip industry is full of twists and turns, and generations of enterprises have gone forward and followed, and under the arrangement of fate and opportunities, the companies that can finally rush out are like morning stars.
With the development of industries such as the Internet of Things, artificial intelligence, and cloud computing, the role of data centers as their infrastructure is becoming more and more important. All parties are also sparing no effort to continue to promote their own development in the data center, as the bearing unit of computing power and data, the server is the core hardware of the data center.
The decline of the RISC architecture, the competition between the X86 architecture and the Arm architecture, the competition between AMD, Nvidia and Intel, the competition between the United States, Europe and Japan are all continuing to be staged, the competition for core chips is becoming more and more fierce, and server chips have become the focus of competition.
Source: The Paper
In the early server field, RISC architecture processor dominates the world, representative products are IBM's Power, Sun's SPARC, DEC's Alpha, MIPS's MIPS and HP's PA-RISC processor, etc., such processors have high performance, good stability, but the cost is too high, so the early servers are relatively high-end, it is difficult to popularize, which also limits the development of the server market to a certain extent.
All of this gradually changed with the introduction of intel's X86 processor with IA architecture, and as the performance of the X86 processor continued to increase, the living space of the RISC processor was still being squeezed.
Unlike Intel, which specializes in producing CPUs, the above-mentioned reduced instruction set CPU development company not only produces CPUs, but also produces its own servers and workstations. But judging from the current development results, Intel, which adopts an open strategy, has come to the end.
Today, we review the development process of the server chip industry, understand the industry pattern and the twists and turns and fates behind the pattern.
Intel's Path to Dominance
There is no doubt that Intel is today's deserved master of the server chip industry.
Modern servers originated with the rise of the tide of the Internet, and the development of the Internet requires a large number of servers to provide support. In the early 1990s, when the Internet first began to develop, Intel, which now dominates the data center market, was just a fledgling in the field of server CPUs at that time, and only slightly modified the Pentium processor used for its Pentium PC as a server CPU.
In 1995, Intel introduced its first processor product, the Pentium Pro, for servers and workstations, marking Intel's official entry into the server chip field. At that time, Intel reached new heights in the design and manufacturing process of the Pentium Pro, revealing its vision for the server market. In 1996, although RISC architecture servers accounted for 78% of the server market, Intel's potential was already beginning to emerge.
Subsequently, Intel launched a number of innovative products such as Pentium Xeon processors on the basis of the previous generation of products, continuously improving its product performance and market positioning. Until 2001, Intel began to release the Xeon series of processors, aiming at high performance, load balancing, high-intensity scientific computing, multiplexed symmetric processing and other features, through architectural innovation and process advancement, continue to seize market share.
At the end of 2018, Intel announced that it will shift from the previous "PC-centric" to "data-centric", which means that the "data center" business will become the main force of Intel, and the Intel data center once accounted for more than half of Intel's total revenue during the transformation period.
Image source: Digital technology explosion
According to IDC's "Global Server Quarterly Tracking Report" data, the revenue generated by X86 servers reached $21.4 billion in the second quarter of 2021, and the revenue from non-X86 servers was about $2.3 billion, accounting for more than 90% of the X86 market. In the X86 market, Intel, as the leader of the X86 camp, has occupied nearly 90% of its server chips.
Intel was able to beat the major RISC processor manufacturers to dominate the data center market, mainly because of the price advantage of X86 servers. The rise of X86 servers has greatly reduced the price of server products, making servers affordable for small businesses and even individual users, which has gradually eroded the market share of RISC vendors, and only IBM can maintain the development of power architecture by relying on the mainframe market.
In addition, wintel alliance established more than 40 years, X86 software ecology is perfect, focus on chip architecture research and development, do not touch the device ecology, and from the cost, performance, ecology of the three aspects, X86 is the most preferred in the early data center, but also the reason for X86's success in the computing market.
The development of server chips
However, long before Intel made its fortunes and even launched server chips, the industry was already full of competition. In a brief review of server history, blogger Yvonne describes decency and decline in the server market for DEC, Sun, IBM and other companies.
DEC and Sun
DEC was the inventor of the minicomputer, and after inventing the minicomputer, DEC became the second largest computer company after IBM for a long time. However, DEC's minicomputers can only provide services in the local area network for a long time, and they are not yet modern servers. It wasn't until DEC developed the RISC chip Alpha 21064 in 1992 and developed AXP servers and workstations based on this chip that DEC entered the server field.
DEC minicomputer (Source: Zhongguancun Online)
Alpha 21064 was very powerful at the time, and the Alpha 21164 chip, published in August 1994, had a processing power of 1 billion instructions per second and was recorded in the Guinness Book of World Records. But because of such powerful performance, DEC's workstation and server prices remain high, and it is inferior to Sun's server competition. However, the later DEC server still did not take the cheap route, and was on the verge of loss for a long time, and finally was acquired by Compaq Computer in 1998, and Compaq Computer was acquired by HP in 2001. HP support for Alpha Servers ended in 2012.
Sun was founded at Stanford University in 1982, and in 1985 Sun invested in the development of its own SPARC streamlined instruction processor, which greatly improved the performance of workstations. Since then, SPARC-architecture servers have long dominated the competition with DEC and HP minicomputers, and in the tide of the Internet in the 90s, Sun's servers have become the first choice of many Internet companies. In addition, Sun also developed a separate Solaris operating system for its own servers, which became a very advanced UNIX operating system at that time.
However, with the bursting of the dot-com bubble in 2001, Sun suffered a sharp decline in sales and a massive preemption of the server market by X86 servers. The beleaguered Sun adopted an open strategy, licensing the SPARC architecture and instruction set to other companies and open source the Solaris operating system.
Fujitsu minicomputer SPARC Enterprise M4000 (Source: Zhihu)
The SPARC instruction set was developed by Fujitsu to produce mainframes and supercomputers, and that's a story for another day.
Mips with twists and turns in fate
Mention of MIPS I believe that everyone is not unfamiliar, today there are still many industrial equipment, routers, memory, switch control chips using MIPS architecture.
MIPS originated from Stanford University's MIPS project and was the first commercial RISC architecture chip to emerge. Friends familiar with the industry should understand that MIPS's business model is similar to today's Arm, which began by taking the road of instruction set and architecture authorization, and then sold to SGI in 1992 for $333 million due to financial problems. After SGI acquired MIPS, it introduced MIPS processors into the IRIS 4D series, and developed many MIPS architecture workstations and servers.
However, after the SGI company in 2006 abandoned the MIPS architecture of the server, MIPS behind the adoption of a large open strategy, many companies are using MIPS architecture to produce embedded chips, before the rise of Arm, MIPS processor was once the world's largest embedded processor architecture, China's Loongson is based on MIPS architecture development.
IBM, which clung to the city
All the RISC instruction set processors mentioned above can still be commercially available on a large scale in the server field, and only IBM's Power is left. IBM was the first company to propose a RISC architecture, but its RISC CPU didn't come out until 1990, later than Sun's SPARC and MIPS.
The popularity of power chips is needless to say, as of now, IBM Power has evolved to the Power 10 architecture after multiple generations of development. Almost all of the Power servers seen on the market so far (mainframes and minicomputers and Power Systems series of servers) are made by IBM.
Compared with Intel's Xeon series, IBM has encountered the same problem as all other declining RISC architectures, that is, the use is too narrow and the scale is too small, resulting in a high unit price of processors, and it is becoming more and more inferior in the competition with Intel.
Probably after Intel's acquisition of Altera, IBM also learned from Sun to engage in the Power Architecture Open Alliance, and established the OpenPOWER Alliance with Xilinx and many device manufacturers, intending to attract partners to build server products under the Power architecture through open source.
Perhaps to emulate the success of Linux, the open source of the Linux operating system has indeed created a miracle around the world, creating a huge ecosystem with the power of a spark that can ignite the wilderness. However, from the current point of view, the Power architecture does not seem to have the same effect after the open source, and the server products based on the Power architecture on the market do not know how many data centers are in use.
Image source: Top net
Taken together, Intel has won an absolute victory in the data center market. Zhihu blogger Yvonne believes that the reason why Intel can gradually enter the server processor market and eventually occupy a monopoly position, in addition to the open strategy of the Wintel alliance, is largely due to the low cost brought by Intel's scale advantage, coupled with Intel's leadership in the process field.
But perhaps the current situation is changing, the rise of Arm processors, TSMC's leading position in process technology, the current lack of only the construction of the ecosystem. Intel was able to defeat a vote RISC server, relying on a cheap X86 server, and today's Arm and Power servers will erode Intel's market in the same way.
At the same time, with the market demand and industry changes, GPUs, AI chips and DPUs have become an important part of the server, and industry manufacturers continue to exert efforts to bring impact to Intel's status.
Intel is ambushed on eight sides
According to TrendForce research, with the growth of market demand, the proportion of server demand in hyperscale data centers is expected to reach nearly 45% this year, from more than 40% in the fourth quarter of last year. Global server shipments are expected to grow by more than 5% in 2021, and companies from all sides are targeting the huge market for data centers.
Market trends are reflected in enterprises, and even the brilliant Intel has ushered in an unspeakable loneliness. In the past period of time, although Intel is still a well-deserved server overlord, it has been caught up and challenged by strong opponents such as AMD, Arm camp and domestic counterparts, and has been frequently questioned by the outside world because of the delay in the progress of advanced processes.
AMD is on the rise
If Intel feels lonely in the server field and wants to find a competitor, I am afraid that it is AMD that can compete with it.
According to the data released by TrendForce, AMD is expected to win 10% of the market share in the server chip market this year, and will win 15% of the market share next year, while Intel has previously occupied 99% of the market share of the server chip market, and the server chip market is a major blow to Intel by AMD.
In 2018, AMD launched the EYPC Dragon server chip, continuing the performance advantages of the Zen architecture, coupled with the price advantage of AMD's server chips, which can be compatible with Intel's X86 architecture server chips, AMD slowly opened up the situation in the server chip market.
Image source: Computer Vision Research Institute
Research firm Mercury Research released its latest server market data for 2021 a few months ago. AMD EPYC Xiaolong series server CPUs reached 8.9% at the end of the quarter, an increase of 1.8% sequentially, the fastest single-quarter growth since Q2 2006, a 15-year high.
However, this is not the peak, before relying on Opteron processors, AMD only took 18 months, the share of the server CPU market increased from 5% to 22%, AMD's highest market share in the server field once reached 27%. However, the market in that year was thrown up in the mountains, and now it is not easy to fight back, and amD's biggest miracle in recent years is actually in the server CPU market, almost from scratch to recover the lost ground.
In March, AMD released the all-new Epyc7003 series of processors, code-named Milan, with the new Zen 3 architecture. It is reported that AMD has completed the design of the chip, using TSMC's 7nm process. Compared with the problem of delayed mass production of Intel's new products, AMD relies on TSMC's strong shipping capabilities to occupy the first-mover advantage.
In addition to maintaining stable performance, xiaolong EPYC processors are more important to keep pace with subsequent product iterations, product operations are steadily moving forward, steady and steady, and the server market also has a world of AMD. Looking ahead, AMD will invest in a 5nm process by the end of 2021, which will further optimize cost, power consumption and performance, and further increase the share of the server chip market.
On the other hand, the news that AMD plans to acquire Xilinx is expected to be a new driver of its development. Nowadays, in the case of high parallelism and high density computing power becoming the mainstream, the traditional general-purpose processor has been unable to meet the demand, and the integrated design capabilities required under this trend must have CPU+ heterogeneous units and ecological collaboration capabilities, which is also the basis for AMD and Xilinx to reach cooperation.
Arm leads the battle for server chip architecture
In addition to AMD, Arm is a big problem for Intel.
There are many products in the server chip market, and the functions and performance positioning are different, but according to the instruction set architecture, there are two main camps in the market at present, one is the CISC (complex instruction set) architecture, and the other is the RISC (simplified instruction set) architecture.
CiSC designers want to make programming easier and programs run faster by building complex instructions directly in hardware, and each instruction in its architecture can perform several low-end operations, such as reading, storing, and computing operations from memory, all in a single instruction; in contrast, the RISC architecture contains only a small number of simple instructions that are used frequently, and provides some of the necessary instructions to support the operating system and high-level languages.
The CISC camp is represented by Intel and AMD's X86 architecture, while the RISC camp includes Arm, MIPS, Power PC, RISC-V and other architectures.
Among them, Arm's development in recent years has been booming.
Arm, the IP licensor that shines in the mobile processor market, gathers the world's top chip designers around, and the shipment volume is tens of billions of yuan to make Intel lose face and fade away. Of course, the saturation of the mobile phone market means that Arm's market share is also close to saturation, and it is necessary to open up new markets to achieve market growth.
Arm's desire to make a difference in the server space is well known, and it has launched an impact on the server market since 2008, but unfortunately the impact has not made many waves. There are many disciples in the Arm camp, but in the field of server chips, it has been very bumpy. In 2008, Arm began to brew a server chip plan, and then riding on the east wind of Arm 64, Qualcomm, Cavium, Broadcom, Microsoft, Huawei, Feiteng, etc. have successively developed their own Arm server CPUs, and later experienced AMD coming and going, Qualcomm high-profile entry but collapsed into the sand, Samsung secretly force is also difficult to escape the end. After more than ten years of development, Arm server chips still cannot have a strong impact on X86, and there are not many companies that still insist on the field of Arm server chips.
However, in the past two years, the Arm camp seems to have begun to slowly become lively again. Ampere, Fujitsu, Mavell, Amazon, Huawei, etc. have frequently launched Arm-based infrastructure processors. In 2019, Arm introduced the Neoverse product line and released the Neoverse N1/E1 core IP – the first time Arm has promoted a "named" processor IP specifically for server products.
Image source: Arm
While arm Neoverse N1 is still not comparable to Intel/AMD in terms of pure performance and ecological integrity, Arm's solutions are attractive enough for infrastructure vendors at the TCO (total cost of ownership) and energy efficiency (Power Efficiency) levels.
Arm-based server chips have been reinforcing their presence in the data center market over the past few years. DigitIMES's research report released in March this year mentioned that cloud computing data center operators such as Amazon and Microsoft, as well as chip design companies such as Marvell and Ampere, have begun to adopt Arm architecture, which is expected to help Arm expand in the large-scale cloud data center and HPC market, or will pose a threat to X86 products.
If Apple released a Mac computer equipped with an M1 processor, so that PC manufacturers feel that Arm personal computers have a head, then the achievements of the Arm server chip camp manufacturers have encouraged more chip designers and provided a new development path.
Despite this, Arm and RISC architecture at this stage only maintain a small number of large-scale order production, and the data center market is the mainstay, TrendForce believes that by 2023, the RISC architecture chip represented by the Arm architecture is still difficult to compete with X86.
Some industry veterans told the author that compared with Intel X86, the gap between other architectures is not performance, but ecology and application, no matter what architecture, and finally promote it or rely on applications, so the architecture is of little significance, to see how to do the application and ecology, to see the activity and support of the architecture ecology.
The prosperity of the ecology must rely on more talent gathering, manufacturer support, and capital investment, which is indispensable. The demise of some structures in history is often not that the technology is inferior to that of people, but that the ecological construction has not kept up with the development of the market. X86 and Arm are relatively leading in talent gathering, vendor support, and capital investment, X86 has a first-mover advantage, while Arm has an advantage in TCO, and in the new market segment, arm architecture can play the unique perspective of each CPU vendor and play its own advantages.
Nvidia wants to break through with AI
NVIDIA is now the most popular in the data center market. According to NVIDIA's 2021 Q4 quarterly report, its data center business revenue reached $1.903 billion, an increase of 97% year-on-year. At present, the data center business accounts for about 40% of NVIDIA's revenue.
In fact, as early as 2011, there was news that Nvidia planned to make its Tegra chip enter the server market, the chip core will set the parallel processing power of Nvidia graphics processor and Arm's low energy consumption characteristics in one, challenging Intel's dominance in the server chip market, but then did not cause much splash, the game business still occupies the majority.
In 2016, with the rise of AI and the beginning of the data center market, Nvidia took advantage of the shareholder wind to ride the momentum, and its GPUs gained key advantages in the field of AI. Nvidia's GPUs have stronger parallel computing, low-precision computing and other performance than Intel's CPUs, and because of this advantage, most of the world currently useSighter's GPUs to train neural networks.
According to a research report by CICC, since 2016, NVIDIA's data center business revenue has climbed all the way, with a year-on-year growth rate of more than 100% for seven consecutive quarters.
NVIDIA's data center business revenue and year-on-year growth rate (Source: CICC Research Department)
Relying on the advantages of GPUs, NVIDIA began to enter the server market, hoping to challenge Intel's monopoly position in the server chip market with the low power consumption and low cost advantages of the Arm architecture.
At this year's GTC conference, Nvidia released the first Arm architecture server processor Grace, for AI, supercomputing, data center and other fields, and will be available in 2023. On the entire computing architecture, the communication protocol between the CPU and the GPU is changed to NVLINK to achieve truly high bandwidth. Grace CPU and computing architecture have the meaning of "revolutionizing" X86 life.
In addition to Intel and AMD's X86 CPUs playing the role of the heart in the server, the AI GPU accelerators installed by Nvidia on the server play the role of training artificial intelligence to enhance inference.
Nvidia's GPUs have a first-mover advantage in the field of AI, most users around the world use their GPUs to train neural networks, and Nvidia lacks a processor platform, which can be integrated with GPUs to complete the delivery of high-performance computing server platforms.
Although Nvidia's acquisition of Arm is still in progress, the ARM-based CPU took the lead in surface, starting with gpus to expand into CPUs. Once Arm is successfully acquired by NVIDIA, NVIDIA's shortcomings in the CPU will also be filled, and then take advantage of the low power consumption and low cost advantages of the Arm architecture. NVIDIA will have full CPU + GPU capabilities, and can build an ecological empire across mobile, PC, and server, which will bring unprecedented pressure to Intel.
Local manufacturers join the battle
In terms of domestic server chips, the instruction set architecture used in the processors being developed in China is all-encompassing, and the most active ones are Loongson with MIPS architecture, Shenwei with Alpha architecture, Feiteng with Arm architecture, and Huawei with Arm architecture.
Although the change has come, but in the server architecture, Intel X86 has firmly occupied the "C" position, domestic manufacturers such as X86, Arm, MIPS, Alpha, Power and other technical routes are different, but are in the performance of server chips to improve the strength, or increase the number of cores, or improve the number of roads, or both, synchronous development, and strive to domestic chips in key industry applications to meet the diverse needs of users.
At present, for the Chinese market, the X86 server business market is huge, including Huawei, Inspur, Haiguang, most of the Chinese server manufacturers, has been unable to get rid of Intel's heavy dependence and restrictions on the underlying chip architecture, there are more risks.
At present, domestic chip companies have made good achievements in arm and alpah two major architecture server chips.
The local server chip manufacturers represented by Feiteng have entered the arm server chip field since 2011, and it has been almost a decade now. With the specific market advantages in China, in the field of domestic substitution, Arm servers have gradually become mainstream.
From a technical point of view, in the server field, especially in cloud computing application scenarios, the low power consumption of Arm server chips is its biggest advantage. Arm server chips through the multi-combination to achieve high performance comparable to X86 chips, and through its low power consumption characteristics, to achieve the purpose of energy savings. Therefore, Arm server chips are a natural pair with cloud, data center virtualization and other application scenarios. At the same time, Arm's ecosystem is also maturing rapidly, especially on the server side, and the progress is very rapid.
Shenwei research and development of Alpha architecture server chip may be China's worthy of the development of server chip direction, alpha architecture is already acquired by HP DEC company research and development, follow-up Alpha architecture was gradually abandoned by HP, after more than 20 years, alpha architecture part of the patent has expired, Shenwei thus the Alpha architecture acquisition into an independent control of the chip architecture. Although the Alpha architecture has almost no sense of existence in the market, but the architecture technology is actually quite advanced, in the 1990s almost along with the X86 architecture, Shenwei bought the Alpha architecture after the launch of the server chip for the world's top level of Taihu Light server, proving the powerful technical performance of the Alpha architecture server chip.
In fact, in addition to the server chips of the above several architectures, China is also developing Power architecture server chips, but under the control of IBM, the performance of domestic Power architecture server chips is slightly backward.
Overall, with the increasing complexity of the international environment and the country's increasing emphasis on independent technology and security control, the localization process of server chips may accelerate development.
Write at the end
Currently, the largest vendor of server processors remains Intel. But perhaps, the AMD, Arm camp and Nvidia discussed in this article will bring strong challenges to Intel. On the one hand, AMD is fierce, and Nvidia has a strong fulcrum of GPU + Arm CPU in the field of high-performance computing, which is expected to slowly encroach on Intel's market from the high-end market. On the other hand, as Internet manufacturers have more and more powerful voices in data centers, Arm-architected server chips may also have an increasing market for terminal demand customers to directly provide customized server chip designs.
After several waves of ups and downs, server chips or will usher in new changes in the near future, opportunities and fate continue to rotate, I don't know who will become the next morning star.
*Disclaimer: This article is original by the author. The content of the article is the author's personal opinion, semiconductor industry observation reprint is only to convey a different point of view, does not mean that semiconductor industry observation endorses or supports the view, if there is any objection, welcome to contact semiconductor industry observation.
Today is the 2803rd content shared by Semiconductor Industry Watch for you, welcome to follow.
★ Chip giants fight each other
★ The Semiconductor Depression is bound to come?
★ Block Samsung Semiconductor
Wafer | integrated circuit | devices | automotive chips | storage | TSMC | AI | packages