Which chips are not easy to be domestically produced?

X86 processor

In 2019, AMD CEO LisaSu confirmed that AMD is no longer licensing its new X86 IP products to Chinese companies (Tianjin Haiguang). It is also worth mentioning that the X86 license obtained by the domestic X86 chip manufacturer Mega Core from VIA has also expired in April 2018. This also means that the prospects for domestic X86 chips have once again fallen into a bleak situation.

Domestic X86 technology source

As we all know, in the PC/server processor market, Intel's X86 architecture is unique in the world with excellent performance, coupled with the deep alliance with Microsoft Windows, making the entire PC application ecosystem built around Intel's X86 architecture. For domestic processor manufacturers, if you want to enter the PC / server market, X86 architecture is undoubtedly the first choice. Especially under the demand background of "domestic independent control", domestic X86 architecture chips also have a "market".

We all know that the X86 architecture is Intel's intellectual property, domestic chip manufacturers to do X86 architecture chips, the first thing that needs to be solved is the problem of X86 technology licensing. For Intel, it is obvious that it will not be willing to license the X86 architecture to Chinese mainland chip manufacturers, fortunately, in addition to Intel, AMD and Taiwan VIA have X86 technology.

AMD

In 1981, as a late entrant to the PC market, in order to quickly launch products and re-establish a technological leading image, IBM used an open architecture for the first time and adopted an outsourcing strategy for the two core components of PCs, the operating system and microprocessors. At that time, the microprocessor manufacturers available for IBM to choose from included at least Motorola, Zilog, National Semiconductor, Fairchild Semiconductor, and AMD in addition to Intel.

Although Intel has a slight upper hand in terms of technical strength, it is still not easy to get absolute support from IBM! Because IBM, which has experienced hundreds of battles, knows that if the microprocessor is completely put to a supplier, it is likely to cause it to be difficult to control, so IBM strongly requires its microprocessor supplier to license the technology to the second supplier, "I am open, you are open"!

The following story has almost no suspense, deep historical roots, many years of cooperation, technical gap is more importantly the blue ocean temptation of the microprocessor market, Intel and AMD quickly hit it off. Intel opened up technology and fully licensed AMD to produce x86 series processors, while AMD abandoned its own competitive products to become Intel's backup supplier. The two sides worked together to finally win the order from IBM, and since then, they have locked the development path of personal computer technology! As many years later, in the lawsuit against Intel, AMD repeatedly emphasized that "AMD's support made Intel immediately change from a chorus member of the semiconductor company to a personal star"!

VIA

As early as 2009, the FTC filed an anti-competitive conduct lawsuit against Intel, accusing Intel of violating the provisions of Chapter V of the U.S. Federal Trade Commission Act of 1914, and arguing that Intel was more extensive than the violation of antitrust law, and that it was an unfair means of competition and deceptive business practices, so it could not be imitated as an antitrust violation through compensation.

Although Intel did not admit it, it eventually reached a settlement with the FTC. Under the settlement, Intel is prohibited from bribing computer manufacturers to purchase only Intel's chips and refuse to purchase chips from other manufacturers, and prohibits malicious retaliation against manufacturers who purchase rival products.

In addition, the settlement requires Intel to amend its intellectual property agreements with AMD, NVIDIA, and VIA to give the above companies greater freedom to merge or joint venture with other companies without being threatened by Intel patent litigation, and Intel also needs to extend the X86 license agreement provided by VIA for five years until April 2018, and retain the PCI-Express bus as a key interface for at least six years. After April 2018, VIA can no longer use Intel's new X86 patents, but the old patents will continue to be used.

In addition, Intel needs to open up to developers the differences between Intel and non-Intel chips in Intel computer compilers, while allowing any software vendor to compile software using non-Intel compilers.

Subsequently, VIA has launched a series of products such as C3, C7, Eden, Nano, etc., but it has always been possible to eat in the embedded market, and the scale is not large.

VIA and Shanghai Mega Core

In April 2013, Shanghai Lianhe Investment Co., Ltd., a subsidiary of shanghai State-owned Assets Supervision and Administration Commission, and VIA Group in Taiwan established Shanghai Zhaoxin Integrated Circuit Co., Ltd., a joint venture, with Chinese state-owned assets occupying a controlling position (about 80% of the shares). As a result, Shanghai Mega Core successfully obtained the X86 architecture license from VIA.

After its establishment, Zhaoxin not only undertook the National Nuclear High-tech No. 1 special project, but also won the support of the Shanghai Municipal Government, becoming a company that did not lack money, and it is said that it received more than 5.6 billion yuan in subsidies from the State-owned Assets Supervision and Administration Commission before and after. Therefore, even if Zhaoxin does not have hematopoietic ability, he does not worry about eating and drinking. The ZX series OF CPUs have also been questioned by the outside world as VIA products wearing vests (for example, ZX-A to ZX-C are the same as VIA's NANO). And VIA itself can't support the wall, and the technology of the processor is far behind Intel and AMD for more than two or three generations.

However, there are also people in the industry, since ZX-D, has no longer been based on VIA products. At present, the latest "KX-6000 series" of Mega Core is said to have comprehensive performance equivalent to Intel's seventh-generation Core i5-7400.

However, with the expiration of via's X86 licensing agreement with Intel in April 2018, VIA has been unable to use Intel's new X86 patents and related software, which also means that M&G's future X86 product upgrades have been blocked, and although it can still use the old X86 patents, the prospects are already bleak.

AMD and Tianjin Haiguang

In 2016, AMD and China's Tianjin Haiguang Advanced Technology Investment Co., Ltd. (THATIC) established a joint venture, Zhongke Haiguang, which can use AMD's X86 technology and SoC IP for chip development. To do this, AMD received $293 million worth of cash, including royalties.

It should be noted that according to Intel's original X86 technology licensing agreement with AMD, AMD has no right to license X86 technology to third parties. However, at that time, a spokesperson for AMD explained the cooperation, saying that the agreement with China did not violate the cross-licensing agreement signed by AMD and Intel in 2009.

Because the ownership structure of the joint venture is different, and all information transferred to China is in compliance with U.S. export regulations. The joint venture company established by Tianjin Haiguang and AMD can indeed modify AMD's CPU core and enjoy X86 authorization in disguise, while Haiguang can develop server CPUs by purchasing CPU cores developed by the joint venture company, but only limited to the Chinese market.

After the establishment of the joint venture, it soon launched a product based on AMD's first-generation Ryzen and EPYC Zen architecture, and successfully applied it to The Dawn server. According to the information exposed in May this year, Haiguang's 32-core, 64-thread x86 architecture server CPU has been successfully streamed. However, the same Haiguang chip is also considered to be an AMD product wearing a vest.

According to AMD Su Zifeng's confirmation at the Taipei Computer Show, AMD's cooperation with Tianjin Haiguang is limited to AMD's first-generation Ryzen and EPYC Zen architectures, and AMD's newly launched Zen 2 microarchitecture design is not available. This also means that Haiguang will not be able to obtain a new X86 patent license and AMD's SoC IP license in the future. This also casts a shadow on the future development of Haiguang, and the gap between it and Intel and AMD in subsequent products will become wider and wider in the future.

From the perspective of the development of Shanghai Zhaoxin and Tianjin Haiguang, a joint venture company has been established with foreign X86 technology manufacturers, the other party provides technology, and we provide the market in order to quickly achieve the "independent and controllable" road, which looks beautiful, but it is actually a pit. And with Arm's entry into the PC and server market, and the rise of RSIC architecture chips, I also see that the X86 unified PC and server market situation is gradually being broken. In addition to X86, domestic chip manufacturers still have other paths to go.

High speed ADC

As we all know, signal chain chips mainly include amplifiers, digital-to-analog conversion, interfaces and other categories, of which converters belong to the highest subdivision categories with the highest technical barriers. The converter is the most critical link in the conversion of analog electromagnetic waves into 0101 bit streams, which can be divided into two categories: ADC and DAC, the role of ADC is to sample the analog signal at high frequency and convert it into a digital signal; the role of the DAC is to modulate the digital signal into an analog signal.

ADCs account for nearly 80% of total demand. ADC/DAC is the crown jewel of the entire analog chip, and the core difficulty is two points: the sampling frequency and sampling accuracy are difficult to combine (the highest barrier of high-speed and high-precision ADCs) and the need for the precision cooperation of the entire manufacturing and research and development process.

Key ADC metrics include "slew rate" and "slew accuracy", with high-speed, high-precision ADCs having the highest barriers. The data converter mainly looks at two basic indicators, the conversion rate and the conversion accuracy.

The conversion rate is usually expressed in units of sps (Samples per Second), that is, the number of samples per second, for example, 1Msps, 1Gsps corresponding to the data converter sampling times of 1 million times and 1 billion times per second, respectively; the conversion accuracy is usually expressed in resolution (bits), and the higher the resolution, the smaller the gap between the converted digital/analog signal and the original signal. High-performance data converters require high-speed or high-precision data conversion capabilities.

For example, Arnold's fastest commercial analog-to-digital converter has a conversion rate of 26Gsps but only 3 bits resolution, while analog-to-digital converters with 24-bit resolution have a conversion rate of only 26Msps.

According to these two indicators, ADCs can be divided into four types: high-speed high precision, low-speed high precision, high-speed low-precision, and low-speed low-precision, of which high-speed and high-precision barriers are the highest.

The current situation of the industry dominated by importers

The industrial chain of ADC chips is the same as that of the semiconductor industry, and its industrial chain is large and complex, which can be divided into: the upstream support industry chain, including semiconductor equipment, materials, and production environment; the midstream core industry chain, including IC design, IC manufacturing, IC packaging and testing; and the downstream demand industry chain, covering industry, communications, consumer electronics, aviation, defense and medical care.

Statistics show that the global converter market size in 2019 is close to 3.6 billion US dollars, and cagr is expected to be nearly 10% in the next 4 years. With the landing of 5G base stations, IoT and other driver ADC demand, it is expected that the global converter chip market space is expected to expand to nearly $5 billion in 2023.

From the perspective of pattern distribution, the global analog chip industry pattern is relatively scattered, and the united States and Europe are in a leading position. Because analog chips have the characteristics of rich categories and deep product series, the scale of each subdivision of products is small, and the span between markets is large, so the global market as a whole presents a decentralized pattern, and it is difficult for head manufacturers to obtain monopoly advantages. The most well-known vendors include ADI, TI, Maxim, Microchip, NXP, Xilinx, STMicroelectronics, and many more. Specifically, European and American manufacturers are still in a leading position in the world due to their early start, with the accumulation of funds, technology, customer resources, brands, etc.

The development of the domestic ADC industry is still catching up. In 1996, 33 countries, mainly in the West, signed the Wassenaar Agreement in Vienna, Austria, which stipulates the scope and scope of export of high-tech products and technologies, of which high-end ADCs belong to export control products, and China is also one of the restricted countries, and the embargo scope is mainly ADCs with an accuracy of more than 8 digits and a speed of more than 10Msps.

After Huawei was included in the entity list in 2019, THE supply of US analog IC manufacturers such as TI and ADI to Huawei was restricted, further accelerating the pace of domestic substitution in the field of domestic analog chips. Focusing on the field of ADC, the world's main suppliers are still several international manufacturers led by TI and ADI, and high-performance ADCs play a vital role in the military field, high-end medical devices and precision measurement, so the domestic substitution of ADC technology is of great significance to the development of China's downstream industries. Since the Huawei incident, domestic equipment manufacturers have gradually begun to purchase domestic ADC chips.

Automotive chips

In this round of chip shortage crisis, China, as the world's largest new car production and sales market, has exposed the long-term dependence on automotive chips, and the "card neck" problem has been further highlighted. Under the crisis, domestic chip manufacturers and car companies have realized the importance of self-developed chips and urgently need to develop high-precision chip technology.

According to iHS statistics, the current global automotive semiconductor market size is about $41 billion, and it will reach $65 billion next year. But in terms of market share, European, American and Japanese companies accounted for 37%, 30% and 25% respectively, while Chinese companies accounted for only 3%. In addition, the data of China's automotive chip industry innovation strategic alliance shows that in 2019, the scale of China's independent automotive chip industry accounted for only 4.5% of the world, the self-research rate of automotive chips in the domestic automotive industry accounted for only 10%, while the import rate of China's automotive chips exceeded 90%, and the domestic automotive chip market was basically monopolized by foreign enterprises.

"The chip problem cannot be solved overnight." Yuan Chengyin, general manager of the China New Energy Vehicle Technology Innovation Center, said that the shortage of automotive chips in China may last for up to 10 years.

At present, the companies with automotive chip design capabilities are some industry giants, such as Intel, Qualcomm, NVIDIA and so on. The only big manufacturers that have the ability to design and manufacture processor chips at the same time are Intel and Samsung.

China's integrated circuit (IC chip) products due to technology, quality and other aspects of there are still many deficiencies, the overall level and the international giants there is a gap of 2-5 generations, but also there is a "short supply" problem, self-sufficiency rate of less than 10%, so long-term high dependence on imports.

Although chips currently account for about 10% of the manufacturing cost of new energy vehicles, with the advent of intelligent and networked vehicles, the cost of chips will continue to increase. It is expected that each vehicle will use 1,000 chips in 2020, so automotive chips are also the key to the transformation and upgrading of the automotive industry.

At present, the demand for automotive chips in China continues to grow, and automobile production and automotive semiconductor composition are two main factors:

In terms of automobile production, data from the China Association of Automobile Manufacturers shows that In 2016, China's automobile production exceeded 29 million, and the global proportion increased steadily;

In terms of automotive semiconductor composition, IHS data shows that the semiconductor composition of each car in China in 2016 was about 235 US dollars, far lower than the level of Japan, the United States and Europe.

At present, as the world's largest single automobile market, China's integrated circuit (IC chip) has been imported for more than 200 billion US dollars for 5 consecutive years, especially in 2017, China's semiconductor chip import cost has been close to twice that of crude oil imports.

The international mainstream chip birth scenario is: the Chinese brand server with the United States eda software design chip, the chip may also use the IP from the British ARM company, and then to Singapore for chip processing and manufacturing, after delivery in Hong Kong, sent to Jiangsu packaging and testing. Most of China's chips were born in a similar process. This crisis has taught us how fragile such processes are. In the above links, in addition to the packaging is still self-sufficient, the top technical capabilities are all stuck in the neck.

Let's take a closer look at the current situation in all aspects of integrated circuits in China:

IC design: The mainland region mainly has the following enterprises in this field: Tsinghua Unigroup, Huawei HiSilicon, ZTE Microelectronics, Huiting Technology, Guoke Micro, Silan Micro, Shanghai Beiling and CLP Huada. In recent years, the growth of China's IC design is still obvious to all, benefiting from the local market, in 2015, China's IC design industry achieved a rapid growth of 26.5%, the scale reached 132.5 billion yuan, and the proportion of China's integrated circuit industry increased from 28.8% in 2012 to 36.7% in 2015. In 2016, China's IC design industry continued to maintain a rapid growth rate of 24.10%, with a scale of 164.430 billion yuan. According to the survey data in 2017, the scale of the IC design industry in mainland China is second only to the United States and Taiwan.

Packaging and testing: The main enterprises in this field in the mainland region are Taiji Industry, Huatian Technology, Tongfu Microelectric, Jingfang Technology, Suzhou Gude, looking at the number of enterprises is not much, but China's mainland region still has a strong strength in semiconductor packaging and testing. In the field of packaging, the technical level and world-class level of China's enterprises have no generation difference, the volume has entered the world's top three, and the development speed is significantly higher than other competitors. In 2012, the revenue of Chinese mainland's integrated circuit packaging and testing industry was only 80.568 billion yuan, and in 2016 it became 152.32 billion yuan, which was 1.89 times that of 2012.

Wafer manufacturing: The three major links of integrated circuits, the mainland region is the weakest in the field of manufacturing. In terms of wafer manufacturing, there are potential stocks such as SMIC, Huahong Semiconductor, Fujian Jinhua and Jinghe Technology in the mainland region. Although the current wafer manufacturing in the mainland region is not strong, the development momentum in recent years is very strong. According to SEMI statistics, from 2017 to 2020, 62 fabs are planned to be built worldwide, of which 26 will be located in China, accounting for 40%.

Looking at the entire semiconductor industry chain, there is another area that we have to pay attention to, that is, semiconductor equipment, wafer manufacturing industry backwardness and equipment backwardness have a great relationship.

Does China really have no chip technology? The answer is no. Pure speed of calculation, we have no problem. After all, in the last global supercomputing competition, the self-developed Shenwei processor won the first place. Cambrian Technology Co., Ltd., a subsidiary of the Chinese Academy of Sciences, released China's self-developed Cambriricon MLU100 cloud intelligent chip, with a theoretical peak speed of 128 trillion fixed-point operations per second, reaching the world's advanced level.

The crux of the matter is that even if you have a chip, even if your chip computes faster in the lab than others. However: what is the computing power of this CPU in the application scenario? Isn't that a big discount? Do you have a chip, do you have a system? Is there an app? Is there ecology?

First, China's chip industry started late, lack of technical reserves, and it is difficult to find relevant high-end technical talents to support research and development. According to data released by the Software and Integrated Circuit Promotion Center of the Ministry of Industry and Information Technology of China, the total number of integrated circuit employees in China in 2017 was less than 300,000, and the gap was 400,000.

Second, Moore's law shows that the chip industry is fast, the investment is higher, and the return is slower. It is difficult for the average enterprise to have strong capital and resource capabilities.

Third, the technical threshold is high. Compared with other consumer electronics chips, automotive chips have higher reliability requirements. The working temperature of general consumer electronic chips is minus 20 degrees Celsius to 70 degrees Celsius, while the working temperature of the vehicle chip must meet the minus 40 degrees Celsius to 85 degrees Celsius, and it must be able to withstand the pressure of hot and cold shock, electromagnetic compatibility, anti-interference and so on. This has formed a certain technical threshold for automotive chip suppliers.

Fourth, the lack of industry synergy mechanism.

With the emergence of new products and new features such as advanced automotive driver assistance systems (ADAS), autonomous driving, internet of vehicles (V2X), and new energy vehicles, technologies such as algorithm chips, millimeter-wave radar, lidar, and new MEMS sensors have developed rapidly. In this context, the integration and mergers and acquisitions of global chip companies are frequent, and the focus of the semiconductor industry has shifted to China.

Graphics chip

As of the fourth quarter of 2020, in the field of integrated GPUs, Intel has occupied 69% of the market share with a stable supply chain, AMD and NVIDIA ranked second and third with 17% and 15% market shares, respectively; in the field of independent GPUs, NVIDIA has an absolute advantage of 82% of the market share, and AMD ranks second with 18% market share.

For domestic GPUs, we have to catch up with the international giants, which have two major directions, one is the GPU chip for graphics processing, manufacturers include Jingjia micro, core technology, Mega core and so on. The other type is GPGPU chips for general computing, and manufacturers include Tianshu Zhixin, Biling Technology, etc. In addition, we also have companies such as Imagination and VeriSilicon in terms of GPU IP. These companies are promoting the development of domestic GPUs in all aspects.

Graphics processing GPU

First look at the graphics processing GPU, the latest generation of Jingjiawei GPU has not been officially announced, but the news said that it will be two JM9 series chips, respectively, using 7nm and 14nm processes, of which the performance of 7nm chip products will reach the level of NVIDIA mid-range GPU.

The GTX1080 compared below is NVIDIA's flagship graphics card released in 2016. Using Pascal, 16nm FinFET process, GeForce GTX1080 has 2560 CUDA processors, core frequency 1607MHz, boost frequency 1733MHz, equivalent memory frequency 10GHz. The graphics card is 256bit wide and has a bandwidth of 320GB/s. The maximum temperature of the public graphics card is 94°C, and it uses a single 8pin power supply with A TDP180w.

As a domestic X86 architecture processor manufacturer, M&G has mastered the three core technologies of central processing unit, graphics processor and chipset, and has the ability of independent design and development of related IP. Public reports say its standalone GPU will use TSMC's 28nm process.

In October last year, they launched a standalone GPU with imagination IMG B series BXT high-performance multi-core graphics processor (GPU) IP, which is a high-performance 4K/8K graphics PCI-E Gen4 GPU discrete graphics chip that will provide strong support for future 5G cloud gaming and data center applications. The process of this chip should be 7nm.

The core technology is also the benchmark Nvidia, the core technology executives said that the current company's GPU architecture is better than Nvidia's Turing architecture, but compared to its newly launched ampere architecture there are still some gaps.

Our domestic graphics processor GPU has made progress with the large-scale shipment of Jingjiawei on the desktop side, and the technology and products of core technology in the data center terminal also have the ability to impact this type of market.

General Purpose Computing GPGUs

Looking at the general computing GPGPU chip, the current progress is relatively fast is the Tianshu Zhixin, in March this year, tianshu Zhixin released a fully self-developed high-performance cloud 7nm chip BI and product card. BI is the first fully self-developed, truly based on the universal GPU architecture of the GPGPU cloud high-end training chip, using the industry's leading 7nm manufacturing process, 2.5D CoWoS package, accommodating 24 billion transistors, supporting FP32, FP/BF16, INT32/16/8 and other multi-precision data hybrid training, integrating 32GB HBM2 memory, storage bandwidth of up to 1.2TB, a single core can perform 147 trillion FP16 calculations per second (147TFLOPS). @FP16）。

BI chips and product cards are released in physical form, and are about to enter mass production and commercial delivery, and the progress of product development and commercial application is 1-2 years ahead of domestic counterparts.

Tianshu Zhixin executives told the electronic enthusiast network that in the wave of AI that in recent years, GPGPU has good support for computing in the form of software written in traditional languages, and has a high degree of flexibility, making it a hot technical product. Compared with ASIC chips, GPGPU has wider applicability, compatibility and flexibility, more prominent tolerance and adaptability to technological changes, and longer application life cycles of products. At the same time, through performance mining optimization, the optimal solution of performance, energy consumption and cost performance is achieved, and the computing power and energy consumption level equivalent to that of ASIC chips are achieved.

Earlier, Zheng Jinshan, chief scientist of Tianshu Zhixin, said, "There is graphic rendering in the GPU of the international leading manufacturer, accounting for about 30% on the chip, but graphic rendering is completely useless for AI and data centers." So we cancel this piece directly. "Maybe that's where the opportunities for GPUs in the data center and AI markets lie."

Founded in 2019, Biling Technology is also a start-up focusing on the research and development of GPGPU chips, committed to developing an original general computing system, establishing an efficient software and hardware platform, and providing integrated solutions in the field of intelligent computing. From the development path, Biling Technology will first focus on cloud general intelligent computing, and gradually catch up with existing solutions in multiple fields such as artificial intelligence training and reasoning, graphic rendering, etc., and achieve a breakthrough in domestic high-end general intelligent computing chips. In just two years, the B round of financing has been completed, with a cumulative financing amount of more than 4.7 billion yuan.

Industry insiders said that the research and development of GPU chips is more difficult than the general chip, the research and development cost is high, and the chip investment cost of advanced technology is huge, as little as hundreds of millions of dollars or more than one billion yuan, so the entire domestic GPU chip is relatively high in both financing and investment. From a technical point of view, gpu design for image rendering is relatively more complex, not only in terms of GPU architecture, but also in terms of interface rate, bandwidth, storage, and advanced packaging such as Chiplet, testing the design capabilities of chip manufacturers. The GPGPU adopts certain computing power and algorithms for AI training or inference scenarios, which can develop computing performance in a more targeted manner.

In the GPU field, it is often said that the success of Nvidia is not only its GPU chip, but more importantly, its software ecology, because of the strong ecology it builds, even if its chips or solutions are more expensive, it is still purchased by users. Such an ecosystem is beyond the reach of many GPU or AI chip companies. Of course, in China, our desktop processor ecosystem is being built, and manufacturers are actively adapting CPUs, GPUs, operating systems and application software extensively. Another opportunity to break through is that domestic GPU chip manufacturers fully communicate with customers from the perspective of customized functions and cost reduction to cut into customer needs.