What are the strategies for the Internet giant BAT in the integrated circuit industry?

In the era of PCs and smart phones, China's general-purpose chip industry is almost blank. Even if domestic mobile phone brands such as Vivo, Xiaomi, etc. come after the end market, mainstream mobile phone chips are still monopolized by foreign companies such as Samsung, Qualcomm and MediaTek.

With the arrival of the AI ​​era, it seems that China's chip industry has brought new opportunities for change. In just a few years, the technology giants represented by the Internet's top 3 BAT have crossed the border, and a number of semiconductor startup companies are also rapidly developing. Rise by the wind.

Below, Xiao Bian has combed the chip projects and companies invested by Internet giant BAT.

Baidu

It can be said that among the three giants, Baidu is the earliest one to devote to the chip industry and operates more systematically. Baidu introduced Stanford University Professor Wu Enda (former head of Googlebrain, who first discovered GPUs to help artificial intelligence implement deep learning algorithms) as Baidu’s chief scientist and responsible for Baidu’s brain research. In addition to developing algorithms, accumulating data, and at the same time trying to develop AI with FPGAs, hardware-related research is involved.

In addition, Baidu also has its own chip R&D project, as follows.

DuerOS Smart Chip

In March 2017, Baidu released its DuerOS smart chip, and reached a strategic cooperation with Ziguang Zengrui, ARM, and Shanghai Hanfeng.

This chip is equipped with DuerOS dialogue artificial intelligence operating system, can give equipment the ability to talk, can be widely used in smart toys, Bluetooth speakers, smart home and other devices.

This cooperation has constructed a three-tier structure including DSM, brain solutions, and chips/modules. Among them, the first two layers are provided by Baidu Ding Mi, which gives the chip DuerOS the “core of dialogue” core functions, including seven major ones. The system 70 has multiple functions, such as schedule management, weather inquiries, answering puzzles, finding music and other artificial intelligence “reservations”, and “unique stunts” such as finding restaurants, ordering meals, and buying movie tickets to directly provide living services. The chip module blocks are jointly supported by Ziguang Sharpui, ARM and Hanfeng.

DuerOS smart chip features "three low, three high" features:

"Three lows" refers to low cost, low threshold, and low consumption. The first is low cost, ie, low price, low development cost, low access cost; the second is low threshold, that is, no need for very complicated professional technology, no need to be proficient in AI domain knowledge and simple code knowledge to apply; Low consumption, high voice compression ratio (1/16), ready-to-use, wide application range.

"Three highs" refers to high security, high integration, and high addedness. First of all, in terms of high security, the smart chip uses ARM's mbed core and security protocol stack to achieve end-to-end secure connection. Secondly, in terms of high integration, it integrates perfectly with the R&D Rising Star RDA5981 chip, MBED, and HF-LPB200U modules. From the WIFI to the narrowband cellular environment can support; the third is high-added, support for voice recognition, voice broadcast, support 100 + artificial intelligence capabilities, support DuerOS brain technology.

DuerOS made its public debut at the CES conference in Las Vegas in January of the same year. DuerOS has reached cooperation with more than 100 manufacturers such as Haier, Midea, Guoan Guangshi and Xiaoyu at home. Expand intelligent voice interaction capabilities to more intelligent hardware, and further work to reduce the development threshold of intelligent hardware.

In addition to the DuerOS smart chip, Baidu will also launch the DuerOS open platform, which will enable partners and customers to customize the DuerOS in the cloud, filter the required services, and provide the ability to connect devices to the cloud and APP.

XPU

In August 2017, at the HotChips conference in California, Baidu announced XPU, a 256-core, FPGA-based, cloud acceleration chip, partnered with Xilinx. Baidu also revealed details on more architectural aspects of the chip at this conference.

In the past few years, Baidu has made great progress in the field of deep learning, especially in the field of deep learning based on GPU. Moreover, Baidu is also developing a new processor called XPU.

The goal of XPU is to balance performance and efficiency and handle diverse computing tasks. FPGA accelerators are very good at handling some computing tasks, but as many small cores are intertwined, the degree of diversity will increase.

Ouyang Jian, a Baidu researcher, said: "FPGAs are highly efficient and can focus on specific computing tasks, but they lack the ability to program. Traditional CPUs are good at general-purpose computing tasks, especially rule-based computing tasks, and are very flexible. GPUs aim at parallel computing. Because of this, it has very strong performance. The XPU is concerned with computationally intensive, rule-based and diverse computing tasks, hoping to improve efficiency and performance, and bring about CPU-like flexibility."

What is lacking in XPU is still programmability, which is also a common problem when it comes to FPGAs. So far, XPU has not provided a compiler. However, Ouyang Jian said that the team will soon develop a compiler.

Figure: The XPU has 256 cores, integrates a shared memory for data synchronization, and all cores run at 600MHz.

Ouyang Jian said: “In Baidu, we have been using FPGAs for many years. We have a large number of FPGAs in our data centers, cloud computing platforms and autopilot projects. We understand the advantages and disadvantages of FPGAs and how to optimize them. With XPU's large core We focus on a variety of computing tasks."

Invest in Lightelligence

In February 2018, Lightelligence, a startup company in Boston, United States, received a $10 million seed round of financing, and the investor was a consortium formed by Baidu Ventures and US semiconductor industry executives.

Lightelligence is speeding up information processing through an emerging Photonic Circuits technology that uses photon ratios rather than electrons to calculate, uses light to transmit signals, is similar to fiber, but is more efficient than electronics, and has lower latency And there is a higher swallow.

The technology was originally developed by scientists at MIT, and Lightelligence is a company that splits the research into commercialization.

The company’s current idea is to use photonic circuits to accelerate the efficiency of artificial intelligence calculations. Several researchers of the company demonstrated the role of photonic circuits in a study published in the Nature-Photonics Journal last year, and in the subsequent series of papers studied photonic circuits and neural network algorithms. Shen Yichen, who was the author of the paper last year, is currently the co-founder and CEO of Lightelligence.

Lightelligence believes that photonic circuits can provide superior performance in many areas, such as coprocessors in the cloud to accelerate the training and reasoning for deep learning. Photonic circuits can also be used in network edge devices to serve systems such as drones and IoT sensors that require high efficiency and low energy consumption.

Lightelligence also stated that there is currently an independent optical chip, and it is calculated that a prototype product will be developed in the first quarter of 2019, and a mass-produced product will be developed in the next 2-3 years.

Alliance with Huawei

In December 2017, Baidu and Huawei jointly announced a comprehensive strategic cooperation. In the future, both parties will launch comprehensive and in-depth cooperation on Internet services and content ecosystems, AI platforms and technologies to jointly build a win-win mobile and AI ecosystem and promote artificial intelligence. The application and full-scene terminal industry rapidly upgraded.

One is ALLINAI's Internet company, and the other is Huawei, which released the world's first AI chip. The strategic cooperation between the two was achieved in the same box of Yu Chengdong and Li Yanhong. For Baidu, it proposed ALLINAI's strategic direction as early as 2016. At the Baidu AI Developer Conference held in May 2017, DuelOS and Apollo were launched at the same time. This made Baidu's AI ecology come into being. .

Huawei not only introduced the world's first mobile-end AI chip, the Unicorn 970, at the IFA 2017 electronic show in Berlin, Germany, on September 2, 2017. It was followed by the release of the HUAWEIMate10, a smart phone equipped with this chip. Moreover, Huawei integrated its resources through resources. It is to build a strategy of coordinated development with core, end, and cloud, and actively deploy AI ecology.

Baidu obtained Huawei's own shortcomings in its hardware through Huawei. It can use Huawei's advantages in AI chips, HUAWEIMate10 and other mobile terminals, allowing its deep learning, voice recognition, image recognition and other AI technologies to get wider and deeper. Landing. Huawei’s use of Baidu, on the other hand, brings AI software level and system level enhancements to its hardware. For example, Baidu’s AI assistant, such as Dudu, may be better integrated on Huawei’s HUAWIMATE series of new mobile phones.

In terms of AI platform and technical cooperation, both parties will use Huawei's HiAI platform and Baidu's PaddlePaddle deep learning framework to leverage the strengths of Huawei NPU (NeuralNetwork Processing Unit) and build an open ecosystem centered on Huawei's AI open platform and Baidu's brain to serve the AI ​​development. It brings richer artificial intelligence applications and smart service experiences to consumers.

Alibaba

As the overlord of China's Internet e-commerce, Alibaba has a mature cloud computing platform, a variety of application scenarios, as well as new retail and strong online sales resources.

Alibaba's platforms generate large amounts of user data every day. These data are used to train AI technology's most fertile nutrients, but at the same time, they also need to have higher computing power, lower power consumption, and smaller AI chips to meet the requirements. .

The AI ​​chip has become a strategic choice. As a giant, Ali will be let go. As former President of the Intel China Research Institute and CEO of Gansu Science and Technology, Wu Gansha said: The success of a chip includes both the success of the technology, the success of the ecology, and the success of the application.

Therefore, in the past few years, Ali has invested in semiconductor companies, accounting for a large proportion of AI chips.

Resistance (Kneron)

In November 2017, Kneron, an artificial intelligence start-up company, announced the completion of A-round financing in excess of US$10 million, led by the Alibaba Entrepreneurs Fund.

Naimeng was established in San Diego, USA in 2014. It has offices in Taipei, Shenzhen, and Zhuhai. The core team members are from Berkeley, MIT and other universities.

The positioning of endurance is a technology provider of terminal artificial intelligence, and now it focuses on lightweight NPU (neural network processing unit) chips, focusing on the terminal market. According to Liu Juncheng, the founder and CEO of Nexus, the company's core competitiveness lies in its lightweight NPU, its power consumption is very low, and its focus is on the end market. The energy consumption ratio can be 100mW to 300mw. The latest product can even go 10mw or less.

Unlike the mainstream cloud-based artificial intelligence currently on the market, Kneron provides an innovative terminal artificial intelligence solution that can transfer a part of artificial intelligence from the cloud to the terminal device for real-time identification and analysis inference, without waiting until all Data can be processed after being transmitted to the cloud via the network to meet the fast and secure requirements, and the burden and cost of the network and the cloud can be greatly reduced.

In 2016, Naimeng launched the company’s first artificial intelligence chip for terminal equipment, called Neural Processing Unit (NPU), and the self-developed software development kit “Reconfigurable Artificial Neural Network”. Using advanced algorithms, you can quickly adjust features for different needs to suit different artificial intelligence applications.

Because the hardware and software can be tightly integrated, compared to the mainstream neural network chips, the capacity needed for the power-resistant NPU can be reduced to 1/40, with better performance and lower power consumption. At present, we have provided customized solutions for smart homes, smart security, and mobile phones. Our related products have been adopted by many well-known manufacturers at home and abroad.

Cambrian

In August 2017, Cambricon Technologies Corporation Limited, which specializes in AI chips, announced the completion of a $100 million Series A round of financing, from SDIC Ventures (A round led castings), Alibaba Ventures, Lenovo Venture Capital, and National Investment Co., Ltd. Zhong Ke Turing, Yuan Yu origin (Angel wheel led cast), Yong Sheng investment (Angel round investor) jointly invested. This made the Cambrian become the first unicorn start-up company in the global AI chip field.

The company's A round of financing was mainly used to promote the product and marketization of the Cambrian series processor in the terminal and the cloud, promote the intelligentization of various types of terminal equipment, and provide cloud-based intelligent processing solutions with high performance and low power consumption. To build a foundation for the smart industry in China and the world, and to build infrastructure.

Cambrian Technology is the world's first successful AI chip company with mature products. It has two product lines: terminal AI processor IP and cloud high-performance AI chip. The Cambrian 1A processor (Cambricon-1A), released in 2016, is the world's first commercially available deep learning processor for smart phones, security surveillance, drones, wearable devices, smart driving and other types of terminal equipment. When running mainstream intelligent algorithms, the performance-to-power ratio exceeds that of traditional processors.

At the end of 2017, Huawei announced the Kirin 970 mobile phone processor, which integrates the relevant Cambrian IP.

Shen Jian Technology

In October 2017, SCI Technology announced that it had completed approximately A$40 million in A+ round financing, led by Ali’s Ant Financial and Samsung Ventures.

Shen Song, founder and CEO of Shenzhen Shenjian Technology Co., Ltd. introduced that part of the current round of financing will continue to be used for product development and market sales in the security and big data field, and more end-products and services for security monitoring and control. On the other hand, the strategic resource injection of Ant Financial will help deepen the further exploration of more application scenarios including finance; and Samsung will focus on cooperation in storage and other aspects to create a deep learning processor for AI chips. As the core of intelligent solutions and efficient overall system, in order to broaden the scope of multi-field products and land, and accelerate the depth of science and technology into the commercialization stage.

Shen Jian Technology was established in March 2016. It was founded by Wang Yu, Han Song, and Yao Song of the Tsinghua Department of Electronic Technology. Afterwards, Dr. Shan Yi, who was in the background of Tsinghua University, was joined as a partner. Shen Jian Technology promotes a deep learning acceleration program, which mainly provides deep compression based on neural network technology and DPU platform to provide an end-to-end solution for deep learning. Through the collaborative optimization of neural network and FPGA, it provides cost-effective embedded and cloud-based The reasoning platform is now used in security and data centers. Shen Jian Technology and FPGA giant Xilinx has always maintained a deep relationship of cooperation.

Barefoot

In November 2016, Barefoot Networks, a software-defined network (SDN) chip company, completed its third round of financing, raising a total of 20 million U.S. dollars. This round of financing was mainly used to open up the market. Internet companies Alibaba and Tencent from China participated in this round. Financing. In June of the same year, Barefoot also received US$57 million led by Goldman Sachs and Google. Together with the funds, Barefoot’s round of round-up financing amounted to US$80 million.

Barefoot Networks has developed the world's first SDN chip. The Tofino chip is faster than any other chip on the market today and handles network packets at 6.5 Tb/s. This type of SDN chip is a revolutionary new innovation that will change the internal operations of Internet companies such as Google and Facebook. It will also affect telecom operators. Hardware giants such as Cisco and Intel will also feel pressure.

Barefoot was established in 2013. The total financing for the three years reached US$155 million. The target customers are cloud computing giants that need to process huge amounts of data. Barefoot co-founder Nick McKeown was the founding father of the P4 programming language. He worked with Princeton University professor Jennifer Rexford on the design and development of P4, and has since pushed forward the development of the P4 community. He also founded Nicira Networks, an SDN company that was acquired by VMware in 2012.

At that time, Cao Jie, an expert on network planning of the Technical Protection Department of Alibaba, once stated that “SDN is a collection of concepts and is our judgment on the evolution direction of the network. We believe that the future network must be application-oriented rather than Today, it relies on switch vendors to dominate the hardware and functions to guide the evolution of the network and the technical direction."

Hangzhou Zhongtian Micro

In January 2016, Alibaba Group’s shareholding in Zhongtian Microsystems became its largest shareholder.

Hangzhou Zhongtian Microsystems Co., Ltd. was established in Hangzhou in 2001. Chief Executive Qi Xiaoning said that from the set-top box to the IC card to the printer, Hangzhou Zhongtian's core business involves many aspects of the embedded system.

Designing a 32-bit high-performance, low-power embedded CPU and developing a chip architecture is the core business of Hangzhou Zhongtian. In 2003, Hangzhou Zhongtian released the first-generation CPUCK510. Since then, the business scope has rapidly expanded to embedded CPUs, SoC platforms, software tools and middleware. At present, Hangzhou Zhongtian has 7 main businesses and 70 agents in China.

At present, the global embedded market seems to have been dominated by two overlords of ARM and MIPS. Why does Hangzhou Zhongtian still have to develop in this field?

Qi Xiaoning said that Hangzhou Zhongtian's ecosystem and tools are not yet comprehensive and mature compared to ARM. "This is our shortcoming," he said, but half of the employees "will continue to focus on R&D chains and simulators." The advantage of Zhongtian is that it can not only compete with others in terms of cost, but also provide “strong support”—helping the new generation of embedded engineers in China, explains Qi Xiaotian. Compared with its competitors, Zhongtian is "more flexible" to cater to consumers' needs.

The trump card of Hangzhou Zhongtian may be the close cooperation with Alibaba. Ma Yun led the Alibaba Group and Tencent, Baidu is known as the Chinese network giant BAT.

Hangzhou Zhongtian is the world’s first chip company to obtain significant investment from Alibaba. Getting investment from Internet giant Alibaba can be quite unusual. For start-up companies, this is equivalent to going from heaven to earth.

Tencent

Compared with Baidu and Ali, Tencent has less attention in the semiconductor field.

Tencent has invested in many artificial intelligence companies, and has established a number of research teams and laboratories. It has also acquired a number of outstanding domestic and foreign companies, but most of them focus on software and applications, involving less research and development of chips.

Barefoot

In November 2016, Barefoot Networks, a software-defined network (SDN) chip company, completed its third round of financing, raising a total of 20 million U.S. dollars. This round of financing was mainly used to open up the market. Internet companies Tencent and Alibaba from China participated in this round. Financing. In June of the same year, Barefoot also received US$57 million led by Goldman Sachs and Google. Together with the funds, Barefoot’s round of round-up financing amounted to US$80 million.

Barefoot Networks has developed the world's first SDN chip. The Tofino chip is faster than any other chip on the market today and handles network packets at 6.5 Tb/s. This type of SDN chip is a revolutionary new innovation that will change the internal operations of Internet companies such as Google and Facebook. It will also affect telecom operators. Hardware giants such as Cisco and Intel will also feel pressure.

The chip will be built into the network switches and hardware devices, and the latter two will play a fundamental role in commanding traffic on the Internet. Switching shuttle data between thousands of computers operated by application manufacturers such as Google and Facebook, and wireless providers such as AT&T, Barefoot Networks chips will change these devices in a significant way.

Its biggest difference with the traditional network chip is its programmable PISA architecture, anyone can program on this chip. The early OpenFlow protocol of SDN required network operators to fill out forms and instruct the switch to handle certain types of traffic. However, PISA does not force converters to use existing protocols such as TCP/IP. Instead, it uses the P4 programming language to allow companies to program switches directly. In other words, they can write software that changes the functionality of the chip, just as anyone can write an application that changes the functionality of the iPhone.

In October of the same year, Tencent senior network architect Lu Sujian said in his speech that Tencent has been doing SDN research since 2012. Currently, the internet export SDN solution has been launched on-line. The investment in Barefoot is also a major move of the two giants in the SDN deployment.

Bit continental

Among current Chinese AI chip manufacturers, Bitland is undoubtedly the hottest company. The main reason is Bitcoin and mining machines, because the company's chips are mainly used in this area.

It is reported that in the Internet field, Bitland has cooperated with Tencent and other giants. However, the specific form of cooperation in investment and financing cannot be determined.

At present, most of the deep learning in the industry is performed in the GPU. Bitcoin selects a dedicated chip and obviously takes another route. In BitContinent's view, dedicated chips are more in line with industry trends.

As the deep learning algorithm matures and the architecture continues to evolve rapidly, the demand for professional chips will exceed GPUs. According to third-party data projections, this turning point will appear in 2020.

Like Bitcoin at the time, it started from the CPU to the GPU to the dedicated chip. Bitland believes that this will take a similar path in deep learning, from the CPU to the GPU to the latest ASIC (TPU). , and think this is the future of deep learning architecture.

It is reported that for the entry into artificial intelligence, Bitland began planning two years ago. Currently, Bitland’s technical reserves in chip design capability, engineering mass production capacity, and artificial intelligence are Bitland’s capital with competitive strength in this field.

At the end of 2017, Bitland launched the TPU chip BM1680, a dedicated custom chip that accelerates processing of tensor calculations for deep learning applications. It is suitable for inference and training of deep neural networks such as CNN, RNN, and DNN (Training).

Bitland has announced plans for the future evolution of AI. At present, the second-generation AI chip, the BM1682 product with 28nm process, is currently being streamed. In September of this year, the 12nm BM1684 will achieve tapeout, and in June next year, it will introduce the fourth. Generation products.

Conclusion

It can be seen that in the treatment of the IC industry, the three BAT strategies are not the same. Baidu insists on self-study and investment walking on two legs. The input to the chip is more systematic and three-dimensional. Ali is a widely used network and has invested in a number of start-up chip companies. Tencent apparently has limited investment in the integrated circuit industry. More energy and capital invested in the software field.