At the end of 2021, Amazon launched its third proprietary processor, Graviton 3, built on the Arm architecture, and applied it to its EC2 cloud servers. This is another breakthrough for Amazon in its self-developed chip development process, and it also puts a perfect end to the surging chip-making process of Internet companies in 2021. This year, whether in the United States or China, the momentum of Internet giants in chip manufacturing has increased.
However, different business environments will lead to different business models. As both are cross-border chip manufacturers, what are the similarities and differences between Chinese and American Internet giants? What impact will their personal participation have on the entire industry?
Common goal: software and hardware integration
The starting point of this chip-making movement was the TPU chip launched by Google at the I/O Developer Conference in June 2015. Google built this ASIC chip to optimize its own TensorFlow machine learning framework, which is mainly used in the AlphaGo system, as well as applications such as Google Maps, Google Photos and Google Translate.
If TPU is just a dedicated chip for AI acceleration, then Graviton, released by Amazon in 2018, is a real "main chip". This chip undertakes the same tasks as Intel and AMD processors in Amazon's servers and performs well, which not only strengthens Amazon's willingness to make chips, but also brings Internet chip manufacturing to a new level.
On this side of the ocean, domestic Internet giants also responded. Baidu announced the development of the first generation of Kunlun chips for AI acceleration in 2018, and mass production of the second generation of Kunlun was achieved in 2021. Alibaba launched the AI chip Hanguang 800 in 2018, and launched the server CPU Yitian 710 using 5nm process in 2021. Tencent announced its three self-developed chips in November 2021: Zixiao, Canghai and Xuanling. So far, BAT has been included in the list of chip manufacturers.
Internet giants in China and the United States have successively invested in chip manufacturing. Is there a common driving force behind this? According to former industry analyst Yao Jiayang, "Whether it is BAT or Amazon or Google, the giants mainly focus on manufacturing mid-range computing chips."
Table: Chip manufacturing by Chinese and American Internet giants (as of the end of 2021)
Elvis Hsu, general manager of the semiconductor division at CINNO Research, believes that the giants have all set their sights on the massive cloud market/cloud server business opportunities, and therefore are developing their own chips that are most suitable for their own cloud computing and AI businesses. Table: Chip manufacturing by Chinese and American Internet giants (as of the end of 2021)
In the past decade, the cloud computing market has exploded at an astonishing rate. The scale and types of business it carries have become quite huge, which has put forward higher requirements for integrated hardware and software systems. If manufacturers want to build data center systems that can meet diverse needs and have high cost-effectiveness, high security and reliability, they must develop chips on their own or work with other chip manufacturers for deep customization to solve compatibility issues in various complex scenarios.
"Self-developed customized AI chips make the performance of cloud servers significantly higher than servers using general-purpose chips such as CPU and GPU. Therefore, self-developed ASIC or FPGA chips has become the best solution, such as Alibaba's Ali-NPU neural network chip," said Elvis Hsu.
Compared with general-purpose chips, self-developed chips can not only meet the specific needs of products and applications, help reduce the energy consumption of equipment and products, but also significantly reduce the cost of servers. It is understood that if self-developed Arm server chips are used, the price is only a fraction of that of X86 chips, or even one-tenth. In an environment where competition in the cloud service market is becoming increasingly fierce, this naturally gives manufacturers great motivation.
Relevant sources revealed that price wars have become a necessary path for cloud service providers to attract new customers and retain old users. The high power efficiency brought by self-developed chips means that lower-priced instances can be launched. Take Amazon's Graviton2 as an example. It beats the performance of AMD EPYC 7571 in almost all workloads, and also beats Intel's Cascade Lake, and the price is more favorable.
In terms of the business path of chip manufacturing, Chinese and American Internet giants also have something in common. After deeply analyzing the chip manufacturing strategies of domestic cloud vendors, Elvis Hsu found that China and the United States have quite a few similarities. In the early stage, both acquired and invested in AI chip design companies, such as Amazon's Annapurna Labs, Alibaba's Zhongtian Micro, Tencent's Suiyuan Technology, etc. Later, they combined the design capabilities of these companies with their own needs and algorithm resources to create exclusive chips.
On the issue of chip manufacturing, Internet giants have basically reached a consensus.
Different considerations: security and ecological expansion
Due to the different environments they are in, there are both similarities and differences in the chip manufacturing practices of Chinese and American Internet giants.
Domestic Internet giants collectively entered the chip manufacturing track around 2018. In addition to being influenced by the chip releases of Google and Amazon, the sanctions on ZTE were also an important trigger.
Although Internet companies are not direct consumers of chips, without the support of chips, all their systems will be castles in the air. Chinese Internet companies need to consider whether they have other options if they face the same dilemma as ZTE.
Tencent founder Ma Huateng once said: "It has woken everyone up. No matter how advanced mobile payment is, without mobile terminals, chips and operating systems, you will not be strong enough to compete."
"One of the main purposes of developing our own chips is to get rid of excessive dependence on chip giants such as NVIDIA, Qualcomm, Intel and other chip manufacturers," Elvis Hsu believes that Chinese and American manufacturers agree on this point, but Chinese manufacturers must also prevent the occurrence of "bottleneck" phenomena.
Internet companies mainly focus on cloud computing, AI and other areas of chip manufacturing, which are the most intensive areas for high-end chip applications and one of the shortcomings of the domestic chip industry. The participation of BAT at this time not only responds to the country's call for the development of hard technology, but also helps to improve the technical level of domestic high-end chips. For example, the Yitian 710 chip released by Alibaba uses the most advanced TSMC 5nm process. Under the ban on HiSilicon, it has objectively become a pathfinder for domestic chips to impact the top process.
If we carefully compare the chips developed by Chinese and American Internet giants, we can find an interesting difference. The types and application scenarios of chips made by Chinese Internet companies are more diverse. For example, Baidu's Honghu chip can be used for in-car voice and smart home, Alibaba's Xuantie 910 is specifically for the AIoT field, and Tencent's Canghai chip can be used for video encoding and decoding.
Industry insiders believe that since BAT's business layout is more extensive and involves multiple tracks, in order to synergize with existing businesses and highlight their business strengths, they will naturally carry out chip layout in areas where they excel.
Like all cross-border players, BAT's chip manufacturing has also been questioned. Some people believe that Internet companies' chip manufacturing is at best customized, and it cannot be considered self-developed. A typical example is that Tencent's Zixiao chip was designed by a chip company it invested in, and the Tencent team was only responsible for the development of software and tool chains.
Most people still hold a more tolerant attitude. Because contemporary chip design is extremely complex, even top companies will not develop 100% of their own chips, and many modules are outsourced or cooperated. For Internet companies, using existing architectures or IPs, combined with their own unique algorithms, to innovate functions can also be considered self-developed.
In the early stages of the chip sector, domestic Internet companies mostly paved the way for investment. Now, whether they are forced by the situation or participate voluntarily, being able to personally enter the game will bring more benefits than disadvantages to the development of the industry.
Opportunities outweigh challenges
As Internet giants begin to make chips, the traditional semiconductor industry has mixed feelings.
Elvis Hsu believes: "Domestic Internet giants such as BAT have entered the ranks of self-developed chips, which will have a certain degree of impact on companies that originally provided general-purpose chips such as Intel, Qualcomm, and Nvidia. These companies have not only lost part of their market share, but may also usher in new competitors."
"On the contrary, it has brought new business opportunities to companies that provide chip design services, such as Broadcom, Marvell, MediaTek, Alchip, etc. In addition, EDA/IP suppliers, leading wafer foundries and high-end packaging plants have also become direct beneficiaries," Elvis Hsu further pointed out.
In recent years, it is precisely because more and more technology companies such as Amazon and Apple have developed their own chips that Intel's stock price has fallen and fallen from the altar. However, the combination of Arm + TSMC has become the biggest beneficiary of this trend and has been promoted to a supernova in the industry. In China, because there are no world-class chip giants, IP and design companies will be the beneficiaries of Internet chip manufacturing, including high-end packaging factories.
Previous article:AMD talks about Intel's big and small core design: Big and small cores are not needed for the time being, Zen has strong scalability
Next article:Latest developments in the UMC Jinhua case: lack of sufficient evidence, Jinhua applies for a not guilty verdict
- Popular Resources
- Popular amplifiers
- Apple faces class action lawsuit from 40 million UK iCloud users, faces $27.6 billion in claims
- Apple and Samsung reportedly failed to develop ultra-thin high-density batteries, iPhone 17 Air and Galaxy S25 Slim phones became thicker
- Micron will appear at the 2024 CIIE, continue to deepen its presence in the Chinese market and lead sustainable development
- Qorvo: Innovative technologies lead the next generation of mobile industry
- BOE exclusively supplies Nubia and Red Magic flagship new products with a new generation of under-screen display technology, leading the industry into the era of true full-screen
- OPPO and Hong Kong Polytechnic University renew cooperation to upgrade innovation research center and expand new boundaries of AI imaging
- Gurman: Vision Pro will upgrade the chip, Apple is also considering launching glasses connected to the iPhone
- OnePlus 13 officially released: the first flagship of the new decade is "Super Pro in every aspect"
- Goodix Technology helps iQOO 13 create a new flagship experience for e-sports performance
- Innolux's intelligent steer-by-wire solution makes cars smarter and safer
- 8051 MCU - Parity Check
- How to efficiently balance the sensitivity of tactile sensing interfaces
- What should I do if the servo motor shakes? What causes the servo motor to shake quickly?
- 【Brushless Motor】Analysis of three-phase BLDC motor and sharing of two popular development boards
- Midea Industrial Technology's subsidiaries Clou Electronics and Hekang New Energy jointly appeared at the Munich Battery Energy Storage Exhibition and Solar Energy Exhibition
- Guoxin Sichen | Application of ferroelectric memory PB85RS2MC in power battery management, with a capacity of 2M
- Analysis of common faults of frequency converter
- In a head-on competition with Qualcomm, what kind of cockpit products has Intel come up with?
- Dalian Rongke's all-vanadium liquid flow battery energy storage equipment industrialization project has entered the sprint stage before production
- Allegro MicroSystems Introduces Advanced Magnetic and Inductive Position Sensing Solutions at Electronica 2024
- Car key in the left hand, liveness detection radar in the right hand, UWB is imperative for cars!
- After a decade of rapid development, domestic CIS has entered the market
- Aegis Dagger Battery + Thor EM-i Super Hybrid, Geely New Energy has thrown out two "king bombs"
- A brief discussion on functional safety - fault, error, and failure
- In the smart car 2.0 cycle, these core industry chains are facing major opportunities!
- The United States and Japan are developing new batteries. CATL faces challenges? How should China's new energy battery industry respond?
- Murata launches high-precision 6-axis inertial sensor for automobiles
- Ford patents pre-charge alarm to help save costs and respond to emergencies
- New real-time microcontroller system from Texas Instruments enables smarter processing in automotive and industrial applications
- Use one data line to control full-color LED lights! Single bus LED usage guide
- 【Portable Programmable Meter】Sliding Rheostat Arrived
- Linux environment programming: from application to kernel
- How to connect the reed switch sensor to the microcontroller
- ssh network login
- What can damage the dp83848 chip? The reset pin is always low and the chip cannot start.
- The 200-meter line signal has irregular jumps after the op amp output
- 【GD32F307E-START】+ Completed driving LCD19264 display
- ESP32 port adds support for ESP32-S3
- TMS320C6000 Pipeline Overview