Last year, we reported that Tencent set up a new company to launch AI chips. A year later, its true colors appeared. After observing and verifying in the semiconductor industry, I learned that Tencent has a team of about 50 people to make chips, and its AI chip has been lost. Today, the AI field has become the commanding height of the world's technology giants. Major cloud vendors have successively handed over their own customized chips, such as Baidu's Kunlun chip, Ali Hanguang, Amazon, Google and Microsoft. What bad intentions can they have? They just want to get cheaper chips or chips with better performance than third parties. Customized chips are getting worse and worse. Who will benefit from it except the chip manufacturers themselves?

The cell-based ASIC business as we know it today was born in the early 1980s and was initiated by companies such as LSI Logic and VLSI technology. Nowadays, this trend is developing more rapidly, and the custom chip market has become popular. Suddenly, anyone with vision and a reasonable budget can make custom chips. Therefore, semiconductor technology is ubiquitous in various customized applications, and products become smaller, smarter and more complex. In particular, it represents cloud computing vendors. Now almost all cloud vendors in the world have entered the ranks of core-building, and they are giving priority to customized design. This is a luxury feast in the chip industry, a feast only for cloud vendors.

In fact, Tencent is a relatively backward member of the domestic BAT nuclear manufacturing ranks. Taking China as an example, Baidu started the FPFA AI accelerator project as early as 20 10, and released the Kunlun chip on 20 18. Now its Kunlun 1 has shipped 20,000 pieces, and Kunlun 2 will also be listed this year. The performance of Honghu chip is not good either. In the past two years, the small size of Honghu chip has occupied the first place in the shipment of smart speakers.

Although Alibaba began to cooperate with Zhongtian Micro to develop cloud chips from 20 15, Ali's core-making wheel went very fast. Acquired Zhongtianwei, merged with Dharma Institute to form Pingtou Ge Semiconductor, and successively handed over Tie Xuan 9 10 and Guang Guang 800 chips to create a full-stack product series integrating terminal and cloud. Moreover, the chip companies it invested in were also very extensive, and almost all AI chip startups were wiped out.

As for Tencent, we all know that it has invested in Suiyuan Technology, an AI chip company, and has invested for four rounds in a row, which shows that it attaches importance to Suiyuan Technology. The performance of Suiyuan Technology is really good. It only took 18 months to complete the research and development, and it was successfully streamed at one time, achieving a breakthrough from 0 to 1, and it was an original chip architecture and an original instruction set. In fact, Alibaba first invested in Zhongtianwei and later acquired it. If this is put on Tencent, Tencent's acquisition of Suiyuan Technology is also a shortcut for chip self-research.

However, the core-building strategy of domestic cloud vendors is still imitating the play of foreign manufacturers such as Amazon to some extent. Let's take a look at the chip research and development ideas of these foreign cloud vendors.

Among foreign cloud vendors, Amazon is at the forefront. In 20 15, Amazon acquired Annapurna Labs, a small chip designer in Israel, and began a long chip journey. Engineers from Amazon and Anna Poorna Laboratories made Arm Graviton processor and Amazon reasoning chip. Graviton chip was originally developed only for special circumstances, but now it can be compared with Intel chip traditionally used in data center, which marks a potential turning point in the industry.

Now, as an important step to control its key technical components, Amazon is developing network chips to power hardware switches that transmit data on the network. It is said that these customized chips can help Amazon improve its internal infrastructure and AWS, and also help it solve bottlenecks and problems in its infrastructure, especially if they also customize the software built on it.

Microsoft created a chip design for the Azure data center and its HoloLens headphones. Recently, it has quietly opened a chip development center in Israel to invest in the research and development of network chips and other products. One of the interesting products developed by Microsoft in Israel is SmartNIC, which is an intelligent network card that can speed up the data transmission in the company's data center servers. The card itself can undertake some necessary tasks, thus reducing the burden on the server CPU. Microsoft currently uses Mellanox's SmartNIC products. But its long-term goal is to replace those products with its own.

Google announced its first customized machine learning chip tensor processing unit (TPU) in 20 16. Google is currently providing third-generation TPU as a cloud service. In recent years, Google has paid more and more attention to chips and hired former Intel executive Uri Frank to lead its custom chip division. Customized chips have always been an indispensable part of Google's strategy of building an efficient computing system. In addition, designing customized server chips will help Google Cloud compete with Microsoft Azure and AWS.

This trend of these technology giants developing their own chips reflects to some extent how different the current technology giants are from the past data center operators. In the past, data center operators did not have the resources to invest hundreds of millions of dollars to design their own chips. Nowadays, the proliferation of custom chips can further reduce the cost of advanced computing products and lead to innovation, which is beneficial to everyone, not only themselves, but also the manufacturers who provide services for them.

Cloud vendors have joined the new ranks of custom chip development, which will generate more business requirements. So, where will all ASIC funds flow? Among them, the obvious benefits are chip design service, EDA/IP demand, OEM demand and so on. Especially those famous manufacturers, but there are also some hidden beneficiaries of the design service industry. In short, manufacturers in these demand areas may benefit from custom chip projects.

First of all, these newcomers lack the accumulation of semiconductor design, which will inevitably lead to trouble. Because there are many chip development processes, including product definition, front-end circuit design, back-end physical realization, manufacturing technology, packaging, etc., and it is often necessary to combine a variety of ip with different functions, which further increases the design difficulty. Not all IC design companies have a deep understanding of these technologies. So we need the help of design service providers.

Among them, companies that provide design services, such as Broadcom, Marvell, MediaTek and Socionext (the merger of Fujitsu and Panasonic's LSI business), will be obvious beneficiaries. Broadcom in particular, it is understood that most well-known manufacturers at home and abroad are using Broadcom's design services. Broadcom occupies a high proportion in global chip design services.

Harlan Sur, an analyst in JPMorgan Chase, pointed out that Broadcom not only helps design chips, but also provides key intellectual property rights for chip production, testing and packaging. Broadcom already has a large number of businesses in the field of network equipment and wireless chips, and may earn as much as $654.38 billion from Google and other companies every year to manufacture customized chips for running servers. Broadcom has been quietly assisting Google TPU in R&D and production. According to foreign media reports, Google's fourth-generation TPU chip has also obtained the service design of Broadcom, and began to design the fifth-generation processor with Google under Alphabet, which will adopt a smaller 5nm transistor design.

Since Google, Broadcom is now helping Facebook, Microsoft, Ericsson, Nokia, Alibaba, SambaNova (a startup founded by Stanford University scholars) and other large companies to provide customized chips that can be used for many purposes.

Marvell's ASIC customization business is also growing. 2065438+In May 2009, Marvell also announced that it had reached an agreement with Singh to acquire Avera Semiconductor, Singh's ASIC business. It is reported that the business department helps chip designers to develop semi-custom chips in fully customized chips. Marvell hopes to shake up the custom ASIC chip market with new 5-nanometer products for 5G operators, cloud data centers, enterprises and automotive applications.

On the other hand, Marvell has been selling a series of chips based on ARM technology for several years, named ThunderX. After selling this chip to Microsoft and other companies for several years, Marvell is now asked to customize a version for Microsoft, which Sur thinks is a cloud computing chip. Microsoft is cooperating with Marvell to develop its next generation ThunderTh3 (TSMC 7 nm) project.

MediaTek began to provide ASIC design services as early as 20 1 1, and has also strengthened ASIC design services in recent years. At the beginning of 20 18, MediaTek officially announced that it would vigorously expand the ASIC design service business, mainly for system manufacturers and IC design companies. The ASIC design service department in MediaTek is an independent department. It is reported that its ASIC design service first favors the Internet giants that have penetrated into the upstream chip design of the industrial chain, and these Internet or terminal giants guide the upstream chip industry and also occupy the bulk of the overall profit of the industrial chain. These are very attractive for the future integrated circuit design service business.

Innovation Electronics (GUC) is a customized IC service provider, backed by TSMC, the largest shareholder, with advanced packaging technology. Innovative electronics is a unique combination of advanced technology, low power consumption and embedded CPU design ability, which is suitable for ASIC design of advanced communication, computing and consumer electronics products. It works closely with TSMC and major packaging and testing companies and has key production technologies.

Alchip is also engaged in ASIC services. According to its introduction in official website, Shixin Electronics can deliver the most advanced ASIC solutions to customers professionally and quickly, and it is the fastest successful manufacturer in node technology such as 16 nm, 12 nm and 7 nm, with a reliable experience record.

Xinyuan Microelectronics (Shanghai) Co., Ltd. (Xinyuan), a listed company of science and technology innovation board, is an enterprise that provides customers with platform, all-round and one-stop chip customization service and semiconductor ip authorization service relying on its own semiconductor IP. According to official website, the core has a rich IP combination in graphics processor, neural network processor and video processor.

MooreElite provides customers with full-process design services from chip definition to product realization, and works closely with many ip suppliers and more than a dozen foundries to create a comprehensive design cloud platform for customers. Based on the long-term polished and verified design process and methodology, Moore Elite has accumulated many years of technology and producible solutions in different application fields such as edge AI, cloud big data training, consumer electronics, industrial systems, network computing SoC and so on.

There are more than 2,000 chip companies in China, most of which are crossing the river by feeling the stones. It can be said that this is a very potential market. According to Gartner, the ASIC market will reach $27 billion in 2020. Dedicated high-end ASIC suppliers will bring huge business opportunities.

In addition to chip design services, most of these chip manufacturers use Arm's ip, and ARM can use its ip to develop customized processors and collect license and royalty income. According to a report by Axios, Google is developing a processor that will power its 202 1 pixel mobile phone and future chromebook. This processor, code-named Whitechapel, is said to have eight ARM CPU cores, adopts Samsung's next-generation 5 nm manufacturing process, and contains special circuits to improve the performance of Google Assistant.

Customized chips need some EDA/IP vendors to provide auxiliary chip design services, and Cadence and Synopsys can undoubtedly benefit from it. As early as 20 18, Cadence cooperated with Google, Microsoft and Amazon to develop cloud-based EDA tools, which can run in the cloud and provide high-level peak performance. Synopsys also cooperates with Google Cloud to widely expand cloud-based functional verification.

And the design and development of all these chips is also good news for the foundry. Chip foundries such as TSMC and Samsung make it easy for technology companies engaged in custom chip projects to obtain cutting-edge manufacturing processes. TSMC is already producing chips, such as Google's TPU and Microsoft's HPU. Last June, 1 1, according to Nikkei Asia, TSMC is cooperating with Google and AMD to develop a new chip packaging technology, 3DFabric, which includes a series of 3D silicon stacking and packaging technologies. It is expected that the first batch of SoIC chiplets will be put into mass production in 2022. TSMC hopes to provide advanced back-end services to its major customers.

Of course, TSMC's move is not an attempt to replace traditional chip packaging manufacturers, but to provide services to high-end customers at the top of the pyramid in order to win over chip developers with strong financial resources. At that time, TSMC won a big order from Apple for its packaging service. Until now, most of TSMC's chip packaging revenue still comes from Apple. Nowadays, in an era when cloud computing tasks are more diverse and demanding than before, customized chips have higher requirements for high-end packaging. If you can provide high-end services for manufacturers such as Google and Amazon, it may be a big income.

Not only cloud vendors, but also Apple's M 1 chip, Tesla's FSD chip and so on. , ASIC chip has become the general trend. Designing ASIC chips requires a lot of capital investment and frequent updates to ensure the adoption of new technologies and manufacturing processes. Technology giants will fight for this technology, and ASIC service providers will not be easy.