AMD has not further authorized chip design for its Chinese joint venture company to prove once again that China's chip development will not be able to take the lead with others' pace.
Text: Bao Yonggang
During Computex 2019, AMD CEO Lisa Su (Su Zifeng) confirmed in an interview that AMD did not further license chip design to its Chinese joint venture. This means that AMD's joint venture processor in China will be limited to the first generation Zen architecture and will not advance to AMD's new Zen 2 architecture.
This proves once again that China's chip development will not be able to take the lead following the pace of others.
AMD initially established a joint venture company, Haiguang IC Design Co., Ltd. with Tianjin Haiguang Advanced Technology Investment Co., Ltd. (THATIC) in 2016. The latter obtained x86 and SoC IP at a price of US$293 million (plus royalties). Authorization for chip development. As China has increased its support in the semiconductor industry in recent years, this partnership has provided AMD with a springboard into the booming Chinese market. The deal is also believed to give China access to key x86 technologies.
Lei Feng.com mentioned in the previous article that some people in the industry have revealed that the joint venture between Haiguang and AMD is a project carried out by some people with the idea of introducing, digesting and re-innovating high-speed rail technology.
According to Tomshardware, the agreement allows Chinese server vendor Hygon to design a dedicated processor based on AMD's Zen microarchitecture, which is the underlying design for AMD Ryzen and EPYC processors. Among them, many of the architecture's customizations include dedicated cryptographic components that meet government requirements. Hygon's first Dhyana processor is very similar to the AMD EPYC data center processor. Tomshardware said they were told that there are other architectural optimizations specifically designed for the Chinese market, but have not yet got more details.
It is not new to establish a joint venture with a Chinese company, but because the x86 architecture license may cause various troubles, AMD's cooperation with Haiguang actually has a complicated design. First, Haiguang parent company Tianjin Haiguang Advanced Technology Investment Co., Ltd. and AMD established Chengdu Haiguang Microelectronics Technology Co., Ltd. (HMC) and Chengdu Haiguang Integrated Circuit Design Co., Ltd. (Hygon). The shares of AMD are 51% and 30% respectively. That is, a relative holding company, and the other is the absolute holding of the Chinese side.
AMD Holdings 51% of HMCs have Zen architecture licenses. In order to remain within the law, HMC licenses IP to Hygon, which designs x86 chips, and then Hygon sells the designs to HMC. But what is confusing is that HMC sold the processor from the foundry to Hygon, which sells the Dhyana processor.
It should be emphasized that these chips must meet the AMD and Intel x86 cross-licensing agreements to avoid x86 authorization leaks and cause opposition from Intel or the United States. In addition, according to the agreement, the final product can only be sold in China.
In the past two years, Haiguang's domestic x86 processor has progressed. Its x86 processor not only has a 4-8 core version of the AM4 slot, but also a server version similar to EPYC. The default output on Twitter is 2.0GHz, the acceleration frequency. Up to 3.0GHz and covers multiple product lines from 4 cores to 32 cores.
Image source: Fast technology
AMD CEO Su Zifeng also said at the Taipei Computer Show that the cooperation with Chinese companies is limited to the first generation of Zen, which means that the follow-up improvements and upgrades of the domestic Haiguang x86 can only rely on themselves.
When asked if he would continue to work with the THATIC joint venture during the US-China trade war, Su Zifeng said that the company is still a joint venture, “we don’t discuss any additional technology transfer” and said that most of the work In joint ventures, there is not much work in AMD.
“THATIC is a single-generation technology license, no additional technology license,” Su Zifeng explained, although she did not clarify whether the decision not to extend technology transfer is a direct result of the trade war. In response to the trade-related impact on AMD's business dealings with Huawei, Su Zifeng said that as a US company, AMD was forced to comply with US regulations.
In addition to AMD, Intel's cooperation with China in chip technology is also relatively limited. In 2016, Tsinghua University, Intel and Qiqi Technology (Shanghai) Co., Ltd. officially signed an agreement in Beijing to announce the joint development of a new general-purpose CPU that integrates reconfigurable computing and Intel x86 architecture technology. Intel will provide funds and other important resources to support project development. .
It is understood that it is possible to develop a coprocessor that connects to the Intel Xeon core to provide China-approved encryption/decryption technology. The two solutions are combined in a package called "Jintide", but Intel does not authorize this to access their core IP.
Tsinghua University, Intel Corporation and Qiqi Technology signed a cooperation agreement
This is another wake-up call. There are two ways for China Core to develop. Previously, Dr. Fang Zhixi, Chairman of the China Advisory Committee of the RISC-V Foundation, said in an exclusive interview with Lei Feng. “There are two roads that prove to be unworkable. One is to close the door and do something yourself. The typical one is the Godson. Because the chip is still a commodity, the performance is high, and no one can use it to reflect the value, so there must be a corresponding ecosystem to play its value. The second way is to follow other people. There are many companies in China that do x86, Arm, and IBM Power chips. In some special fields, these instruction set architectures can actually do some things, but because of the intellectual property rights of the companies affiliated with ISA ( The control of IP) is difficult to achieve success. ”
Is the AI chip an opportunity for the Chinese core?
In 2012, Jeff · Geoffrey Everest Hinton published an article on NIPS that introduced the large error rate by classifying images in the ILSVRC training set by training large deep convolutional neural networks. Lower, this is also considered the beginning of this round of AI boom. Subsequently, deep neural networks became increasingly popular, and many excellent variants emerged, such as AlexNet, GoogLeNet, VGG Net, and ResNet.
Of course, the popularity of deep neural networks is inseparable from the support of computing power. Since most of the operations in deep neural networks are matrix operations, GPUs that are good at parallel computing are well suited for deep learning. NVIDIA quickly seized the opportunity to launch the first Pascal GPU optimized for deep learning in 2016, which also enabled Invitro to reach the most profitable AI chip company.
In fact, before NVIDIA launched a GPU optimized for deep learning, AI chip-related research began to become a hot topic in academia and industry in 2015. Horizon, Endurance, Shenjian Technology, Exploration Technology, Beijing Qingwei Electronic and other AI chip startups have been established. Yuntian Lifei, Bitland, Hangzhou Guoxin, Ruixinwei, etc. are AI chip companies established before 2015.
However, 2015 is only the beginning of the AI chip boom.
AI chip boom meets capital winter and industry recession
In the AI boom, NVIDIA and Intel have increased their AI chip orders and increased their investment in AI. It is also worth noting that Internet giants who are known for their software have also entered the AI chip field, including Google, Facebook, Amazon, Ali, and Baidu.
In addition to the giants, capital is also driving the AI chip boom. Beginning in May 2018, Probing Technology, Yunzhisheng, Nenergy, Spirit, and Cambrian have announced new rounds of financing. During the Spring Festival of 2019, the report said that the Adobe Group is considering whether to build its own chips. At the end of February, Horizon announced that it had completed nearly $600 million in Series B financing, with a post-investment valuation of $3 billion.
The boost of capital and the entry of head companies have caused some people to worry about the overheating of the AI chip market, but the views of the industry are mixed. In January this year, Horizon's co-founder and vice president of algorithms, Huang Chang, said in an interview with Lei Feng. The AI chip's fiery degree in 2018 was not unexpected.
Su Dong, vice president of Walden International, said in an interview at the beginning of the year that the AI chip is not overheated. The future will not overheat or that the most important thing about the bubble is whether the application of the AI chip can be put to the ground and will not overheat. In the cold winter of capital, investment institutions are still investing in head enterprises for security reasons. For non-head enterprises, as investors' requirements for investment targets will be higher, it will be easier to carry out the next round of financing because they focus on faster cash flow.
However, the cold winter from the second half of 2018 has put a lot of pressure on the chip of AI chip startups. The International Semiconductor Industry Association (SEMI) previously lowered its global semiconductor equipment outlook for 2019, and expects its full-year output to decline by 9% from last year, ending its three-year growth trend since 2016.
The latest data from several analysts also demonstrate the challenges facing the semiconductor market this year. On May 7, IHS Markit pointed out that the global semiconductor market situation has deteriorated rapidly at the beginning of this year. It is estimated that the global semiconductor market sales will decrease by 7.4% year-on-year to US$466.2 billion, which may hit 2009 (Note: the year fell nearly 11) The largest decline since %). On May 15th, IDC's latest report pointed out that the semiconductor industry is struggling to cope with the slowdown in smartphone sales and weak demand in China. After three consecutive years of growth in global semiconductor revenue, it is expected to decrease by 7.2% in 2019.
The landing of the AI chip has started
Under the rigorous market background, especially after the new round of AI boom has lasted for four years, the landing of AI chips is particularly important. Su Dong is optimistic about the landing of AI chips in driverless and security. He said: “There will be more applications in the security field in 2019. Whether using Hisilicon chips or related algorithm companies’ chips, it is foreseeable that there will be more More landing applications. Driverlessness is also a good landing scene for AI chips, but unmanned driving belongs to the front-loading market of automobiles, and the cycle will be relatively long. ”
Huang Chang believes that new killer apps may come from smart driving and smart IoT scenarios. It is reported that the horizon began to invest in AI chip development more than three years ago, and the first generation of AI chips has begun to be commercialized on a large scale. In the field of intelligent driving, Horizon has established partnerships with Tier1s and OEMs in the world's four major automotive markets; in the AIoT field, the Horizon has already reached cooperation with modern shopping malls, home life plazas, and well-known brand stores.
With the advantages of algorithms and industry applications, Unlimited will master the core chip technology as the next step strategy, and gradually increase investment in chips while advancing the front-end intelligent sensing products of the security industry, intelligent box/analysis host and building solutions. . In June 2018, Inspirit IoT, an innovative IoT chip design company, set up a strategic alliance for sensory chips and strategic cooperation. Inspirit IoT proposes an optimization method that is 10 times higher than the current industry's compression ratio, which will greatly improve the efficiency of FPGA use. This method will also play an important role in the application of FPGA in the Vista Extreme Plan. In addition, Infinity will also join hands with well-known domestic chip companies to create a new generation of artificial intelligence sensing chip (SOC), providing end-to-end comprehensive solutions for the next generation of artificial intelligence perception.
There has been a phenomenon-level product, Yuntian, “Deep-sighted”, and Yuntian Lifei, a dynamic portrait recognition system, also released the DeepEye1000 embedded visual AI chip with independent intellectual property rights in October last year, which can provide processor-level instruction set flexibility. Software programmable, widely used in cameras, robots, drones, and AI edge computing scenarios such as smart cities, smart communities, smart manufacturing, and new retail. According to Li Aijun, vice president of Yuntian Lifei, the AI chip will be officially commercialized in 2019.
Lu Yong, founder and CEO of Probing Technology, said in an interview that the previous technology did not reach a relatively ideal state, and the cost did not fall to a level suitable for popularization. In 2019, technology and cost will usher in an inflection point. Based on the increasing acceptance of smart homes by users, smart homes will also become popular in 2019, which will also drive the landing of AI chips.
Recently, Liu Juncheng, the founder and CEO of Nike, said in an interview: “The competition for AI chips has reached the competition of application landing. ”
The chip is a long-cycle product. Four years later, we saw that the chips of the startups have been streamed one after another, which also indicates that the landing of the AI chip has started.
The key to the successful landing of AI chips
For AI chip startups, whether the AI chips invested heavily and spent years of research and development can win the favor of customers, continuous iteration of revenue is the key to the company's survival.
For the giants, the strength of the AI chip will determine whether they can continue to lead the AI revolution. Intel is the chip leader. Although its CPU is not as good as the GPU in the AI era, they are enhancing the strength of AI through innovation and the development of independent GPUs.
At the 2019 Intel China Media Sharing Conference last month, Song Jiqiang, president of Intel China Research Institute, pointed out that any single factor is not enough to meet the diversified computing needs. Looking to the future, Intel is driving computing innovation in all directions, leading the era of “hyperisomeric computing”.
He revealed that Intel's architectural innovation as the main driver of innovation in the next decade will continue to bring exponential expansion. The architecture mentioned here refers not only to the design of the CPU core, but also to the composition of the entire computing product from hardware to software. This is because soft and new hardware can achieve more than one hundred times acceleration through software optimization, which better meets the computing needs of AI. For example, after Skylake optimizes through hardware and software collection, AI reasoning performance can be improved by 275 times.
One of the keys to architectural innovation is the breakthrough of the von Neumann architecture. In the AI chip implementation, based on the von · Neumann architecture, providing computing power is relatively simple and easy, but due to the speed difference between the computing component and the storage component, when the computing power reaches a certain level, the speed of accessing the memory cannot keep up. The speed at which the computing component consumes data, and the increase in the number of computing components, cannot be fully utilized, that is, the so-called "memory wall" problem arises.
At present, many startups and teams are innovating in the architecture, using methods such as reconfigurability, reorganization, integration, and storage priority.
Of course, technology leadership is only one of the keys to the AI chip landing. An AI chip has better performance and leads the technology. It can't be favored by users and can't be successful. This involves the cutting-edge field, business model and market environment. The influence of other factors.
On July 13, 2019, the CCF-GAIR Global Artificial Intelligence and Robotics Summit, hosted by the China Computer Society and co-hosted by Lei Feng.com and the Chinese University of Hong Kong (Shenzhen), will open the “AI Chip” special event to “Cool Winter”. Under the theme of “breaking through the AI chip landing problem”, the industry's leading scholars and AI chips are invited to share their views on the future evolution direction of the AI chip architecture, and how to choose the right application field for the AI chip with technical advantages. The unique insights of landing.
At present, Intel's chip experts, Horizon co-founder, algorithm vice president Huang Chang, Yun Tianlifei vice president Li Aijun, Touch View Unlimited co-founder and CEO Xiao Hongbo have confirmed attending the AI chip forum and giving a speech. Professor Wei Shaojun, the chairman of the IC Design Branch of the China Semiconductor Industry Association and the director of the Microelectronics Institute of Tsinghua University, will also be confirmed to attend the conference.
As a new instruction set architecture that is currently attracting attention, RISC-V also has great potential in the field of IoT and AI. In this AI chip special, the researcher of the Chinese Academy of Sciences, the director of the Advanced Computer System Research Center, and the China Open Instruction Ecology ( Dr. Bao Yungang, Secretary General of the RISC-V Alliance, will also share the advantages of the RISC-V instruction set architecture and opportunities in the AI and IoT era.
Lei Feng network reference tomshardware, fast technology