One or two ago, the news that Ali , a chip company under Ali , was widely circulated in the chip industry. Recently, there is news that the flat-headed brother is no longer independent. Behind the repeated information, how are the chip business of major Internet manufacturers doing?
article｜Zhao Yanqiu
editor｜Zhou Luping
One or two ago, the news that Alibaba's chip company Pingtou Ge was going to be independent from Alibaba Group was widely circulated in the chip industry. Recently, there is news that the flat-headed brother is no longer independent.
Alibaba is a representative enterprise in the core building wave of Internet giants. Since 2018, Baidu , Alibaba, Tencent and ByteDance have successively announced their chip business. Four years have passed, and for complex chips, these major manufacturers are still in the early stage of twists and turns. However, industry insiders believe that the speed of large factories is not slow, but they have to go through the hard work of building an ecosystem.
01 Kunlun core can be independent, why does the flat-headed brother repeat?
rumor that the flat-headed brother is independent to a certain extent related to team mobility, and some people have left their large chip teams such as server chips. The industry usually calls high-end complex chips such as CPU and GPU large chips. The market was very hot last year and the talents were in short supply.
"Ali's stock dropped a lot last year. In contrast, a number of external chip companies have successfully listed, and the stocks have risen very high. Some people inside may have ideas." Chip veteran Li Yi told the frontline of the Digital Intelligence, "The key to this matter is the difference in PE between the chip business ( price-to-earnings ratio ), which is purely related to the capital popularity."
According to public information, in the past two years, the price-to-earnings ratio of some listed chip design companies has reached more than 100 times, while Alibaba's price-to-earnings ratio once fell to about 15 times, forming an inverted reversal.
Therefore, while the flat-headed brother is independent, there are also speculations about the flat-headed brother's valuation, with the highest reaching US$10 billion, which is "quite amazing", and some valuations are also said to be one or two hundred billion yuan.
"But the chip market has become popular recently. Some good chip design companies have a price-to-earnings ratio that is about 30 times, and they are all more than 30 times. It may be a little more realistic." The above-mentioned person further said that there may be rumors that there are no independent ones.
One year ago, Kunlun Xin announced its independence from Baidu and has conducted a round of financing with a valuation of 13 billion yuan. This also makes the "independence" of flat-headed brother seem more natural.
"Ali Flat Head Brother is different from Baidu's Kunlun Chip." Another veteran Liu Zhuo told Digital Intelligence Frontline that Baidu's current size is quite different from Alibaba and Tencent, which also leads to a considerable gap between Baidu's data center scale and artificial intelligence chip and GPU chip usage. Therefore, if Baidu's self-developed Kunlun chips are used internally, it may reach a relative bottleneck period, and it needs to be exported to the outside world to have economic benefits.
Another senior person engaged in artificial intelligence chips, Zhang Wei, also told Digital Intelligence Frontline that as far as he knows, Baidu has great support for Kunlun Chip, and the largest search business and Xiaodu have used Kunlun Chip. But he observed that the internal usage of Kunlun chips was close to the magnitude of Baidu's purchase of GPU chips. Kunlun Chip is mainly used in artificial intelligence scenarios, similar to GPU chips.
"And the artificial intelligence chip is updated with the server. Now the server update cycle has been extended from the past 3 years to 5 years." Zhang Wei said, "Baidu's internal demand is no longer enough to support its development." One of the main intentions of the self-developed chips by major Internet manufacturers is to reduce costs. They have their own huge data centers and purchase a lot of chips in a year. Customized development of some chips according to their own business needs not only improve efficiency, but also saves a lot of money.
"But if the money I saved is not only not profitable, but may still lose, I must export it to the outside world to maximize the return on investment." Liu Zhuo explained, "And in the future, I can make another wave of money. So, I guess Alibaba also has this idea to maximize the interests of the flat-headed brother."
02 The core manufacturing of major manufacturers is still in the early stage
Internet manufacturers began to build cores in an established manner, starting from Baidu's release of Kunlun chips in July 2018. Immediately afterwards, Alibaba announced the establishment of Pingtou Ge in September of that year, and Tencent and ByteDance also launched their own chip businesses in the following years.
The incident of ByteDance's core manufacturing was revealed in March 2021. Byte later confirmed that a chip team is being formed. "Byte's overall style is to sneak into the village and shoot. I feel that they have poached many people from major companies, but there is little public publicity. This is also related to the style of Zhang Yiming ." Liu Zhuo told Digital Intelligence Front. According to his speculation, the Byte Chip team may have a size of hundreds to thousands of people.
And people close to bytes told Digital Frontline that not long ago, a chip of Bytes had been cut (trial production). According to previous public information, Byte independently developed cloud artificial intelligence chips and Arm server chips.
In addition to bytes, Baidu Kunlun Chip officially announced the launch of Kunlun second-generation artificial intelligence chip in August of that year after its independence in June 2021.
"'s artificial intelligence chip of Kunlun is put on the open market, and its application field is directly related to Nvidia's GPU." Zhang Wei, a senior person, told Digital Intelligence Frontline. Nvidia is currently the world's largest chip company with market value and the largest GPU supplier, and GPU is the chip with the most widely used in the field of artificial intelligence. But in Zhang Wei's view, Kunlun Chip faces challenges in the open market without Baidu's parent company.
The biggest challenge is R&D resources and rhythm. "Artificial intelligence is a high-tempo industry, with new networks coming out every year. In this market, Nvidia launches a new chip architecture every year." He explained that if Kunlun Chip is at the current pace of almost two years, the gap with Nvidia will widen. It is reported that the new generation of Kunlun chip will be launched in 2022.
This is not a problem faced by Baidu Kunlunxin. Liu Zhuo noticed that there is a characteristic of "demolition of large ships to build sampan ". In the past two years, due to the popularity of in the capital market, many small teams have emerged. The R&D resources of these small teams are not as good as Nvidia, which has a very rapid advancement speed, and "the gap is widening."
In addition, for Kunlun Chip, the open market is not as high as the internal tolerance, and can carefully polish the product. "They need endurance and tenacity," said a chip veteran close to Baidu.
Digital Intelligence Front learned that the total amount of Kunlun second-generation chip orders has exceeded 100 million yuan, including the local intelligent computing center, courts, judicial prosecution, scientific research platforms and other projects acquired by the company.
Some industry insiders believe that Kunlun Chip should try its best to implement the biggest features of the chip, such as the characteristic technology of Baidu Fengchao advertising recommendation system, to the second- and third-tier Internet commercial markets, and achieve long-term sustainability.
Alibaba Flat Head Brother is the most eye-catching among the core making trends of several Internet manufacturers. It first launched the artificial intelligence chip Hanguang 800. In the industry's view, Hanguang 800 shoulders the dual tasks of external publicity and internal development. "The indicators are very high", "there are a batch of internal requirements, and the process of application running is quite painful", which is not considered successful in the eyes of many industry insiders. After
, executives who airborne from HiSilicon are more optimistic about GPUs in cloud artificial intelligence, so they adjust the technical route of Pingtouge's artificial intelligence chips, no longer make special AI chips similar to , 800, and instead design general computing GPUs.At present, this chip has been cast (design completed and manufactured), and the industry rumored performance is higher than that of the Nvidia A100. This chip is expected to be released in this fall Yunqi Conference.
In addition to artificial intelligence chips, Alibaba also announced another big chip last year - the server CPU Yitian 710. "The server CPU is the last vacant piece in China, and no major manufacturers have come to the forefront." Li Yi told the Digital Intelligence Front. At present, Yitian 710 is used in Alibaba Cloud business, and other internal applications are also discussing the use of this chip.
In addition to the above-mentioned large chips, Alibaba Pingtou Ge also has a small chip business RISC-V chip business, focusing on the edge side and terminals. Alibaba's Tmall Ghost plays an important role in the promotion of this business.
However, outside world observes that the "Buddhist" Tencent has been a bit shaky in chip research and development, and in the early days, it mainly adopted the investment route in the chip field. For example, in August 2018, Tencent led the Pre-A round of domestic artificial intelligence chip startup Suiyuan Technology.
Liu Zhuo believes that one of the major backgrounds for Tencent to adopt the investment route is that its artificial intelligence usage is not as large as Alibaba, and the artificial intelligence in game scenarios is relatively complex, and it tends to reinforcement learning and confrontation learning in deep learning. "These strategy-oriented artificial intelligence is more difficult. Compared with artificial intelligence in e-commerce and other scenarios, I think it is difficult to do."
Another senior person close to Tencent told Digital Intelligence Frontline that Tencent's core business, such as games, runs stably, and cost issues are not the first factor that the team considers. The team prefers to choose new functions, so it does not have strong internal driving force for self-developed chips. Some people inside Tencent also tend to cooperate with external forces.
Digital Intelligence Front learned that Tencent AI chip Zixiao and DPU chips were promoted by Tencent and its investment companies such as Suiyuan Technology and Yunbao Intelligent. Among them, chip design is done externally, and Tencent is responsible for software and tool chain development internally.
In November 2021, at the Tencent Digital Ecology Conference, Tencent officially announced the self-developed artificial intelligence chip Zixiao, video transcoding chip Canghai, and intelligent network card chip (DPU) Xuanling. Among them, Zixiao has been streamed and successfully lit up (it is available, and can be mass-produced after subsequent tests and modifications). The DPU chip Xuanling is planned to be launched at the end of 2022.
Some people believe that in terms of chip business, domestic Internet manufacturers are slower than overseas Google and Amazon . But Liu Zhuo believes that "the country just started late, not slow." Moreover, Amazon used to make artificial intelligence chips through acquisition teams, while several domestic companies basically recruited their own troops.
and Zhang Ting believes that the growth rate of domestic Internet companies has been basically the same from the previous 50% to now, and the chip business must be different from the expectations at the start-up, and it has been more or less affected.
03 Building an ecosystem is a more difficult battle
Whether domestic or overseas, major Internet companies cannot escape the chip manufacturing categories: cloud artificial intelligence chips (AI chips), CPUs and GPUs. In addition to being closely related to its own business,
was not only closely related to its own business, but Liu Zhuo told the Digital Intelligence Front that the downstream industry driven by chips is also a huge market.
Take Nvidia, the largest player in the GPU market, as an example. Last year, the global revenue was nearly US$27 billion, and the Chinese market accounted for 1/4, about US$7 billion. Nearly half of them belong to GPUs, which is $3 billion. Everyone is grabbing this cake.
"This is just a GPU card. The server market built with GPUs may have to be enlarged by two to three times, while data centers cannot be just servers, as well as storage, network, operation and maintenance... In the end, the comprehensive market may be 5 to 10 times that of chips." Therefore, people see that large companies are not just making chips, they also build servers, storage, energy modules, data centers...
. According to relevant reports, some investors believe that if a GPU company can cut 5% to 10% from the Chinese market, it will be enough to support a market value of 100 billion yuan.After this account, it is no surprise that big factories are involved in the competition.
However, if Internet giants want to compete with traditional leading chip giants, the most difficult thing is probably the ecological battle.
Taking GPU as an example, Nvidia took 10 years to establish the CUDA ecosystem and received support from a number of software and application companies. Because of this ecosystem, commercial companies generally dare not replace Nvidia. Liu Zhuo revealed that even if other GPU cards are 1/3 cheaper than Nvidia's cards with the same performance, it is useless, because it is too expensive for customers to migrate the software originally running on Nvidia's GPU to a new platform.
"This is similar to what I originally said in English, but now I have to change to Spanish ." Liu Zhuo said.
Previously, another GPU supplier AMD claimed that it could be 99% compatible with the Nvidia CUDA ecosystem, which means that customers can use it without migration, which seems to be a feasible way. But in fact, AMD's GPUs are rarely used in AI facilities in data centers. Liu Zhuo analyzed that CUDA is naturally optimized for NVIDIA, and even if AMD is compatible with it, it is difficult to perform its own performance. Another key point that has been ignored is that scientists or developers around the world are doing original innovation and development on Nvidia GPUs. "It has a continuous stream of ecology and innovation." This makes everyone inseparable.
Among domestic companies, Huawei was the first to build an ecosystem of artificial intelligence chips. "Look at Huawei's AI chip Ascend, there are teams of thousands of people promoting it in the ecosystem, and dozens of universities investing hundreds of millions of millions of dollars. It is actually quite difficult." A veteran said, "To be honest, I think after this wave of artificial intelligence, Huawei may find that it is underpaid, unless this wave of artificial intelligence can continue to expand the market by 10 times or 50 times." After Alibaba's GPU comes out, it also faces the same problem. The most realistic way now seems to be to rely on information innovation markets, and one or two companies can grow from these companies that make chips bigger.
Of course, senior chip expert Wu Wei also saw some optimistic information. "For example, in the CPU field, it turned out to be the world of Intel . In recent years, Arm has developed into an Arm server CPU, and it has caught up with a good opportunity." He gave an example that overseas, Amazon Arm server CPU has been applied. Although many software only supports Intel's ecosystem, the process of ecological migration has begun.
Li Yi and Liu Zhuo also noticed that Ampere, an Arm server chip startup founded by former Intel president Renee James, has also opened up the commercial market. Not only in the United States, Ampere's chips have also been used in domestic public cloud service providers. Its valuation has reached $6 billion and has secretly filed IPO documents to US Securities and Exchange Commission (SEC). "And some domestic companies believe that they are no worse than Ampere," said Li Yi.
Among domestic companies, who is possible to build a big chip ecosystem?
now seems a bit unsolvable. In terms of chip investment and execution, Alibaba's flat-headed guy ranks first, Huawei retains its large chip business, but the advanced processes are limited, and many startups or traditional companies do not have enough scale and resources. "Ali's execution is definitely among the top three in the country," said Liu Zhuo.
However, investment in the ecosystem depends on Alibaba's support for the chip business. If it values chips, it will mobilize Alibaba's internal resources to invest. Last year, Alibaba made a low-key organizational adjustment. Pingtouge was on par with Alibaba Cloud Intelligent and DAMO Academy departments, and is no longer just a department under the DAMO Academy.
"The big chip business is very expensive, and there is no output in the first five years. Therefore, there is enough capital to do it. The reason why Huawei HiSilicon is successful is that it has been continuously spending money in it." Li Yi, who has been engaged in chip business for many years and has experienced successful listing of the company, concluded, "The other is that you can hold on in the middle stage and have money to keep coming in. Anyway, chip design companies will definitely be clear in 10 years.”
