AI chip field BAT or investment or self-study, frequent action
At the time when ZTE was sanctioned by the US Department of Commerce, and public opinion fermented, Ali announced on Thursday that its newly established research and development organization - Exploration Adventure Momentum and Prospect Academy - will work on the development of an artificial intelligence chip called Ali-NPU. The chip can be used by anyone through public cloud services. The chip will be used in imagery video analysis, machine learning and other AI reasoning calculations.
Ali's investment in the chip field adheres to the principle of wide networking, including the investment in Zhongtianwei, the Cambrian, and endurance. The direction of development is ASICs or NPUs that foreign manufacturers have not monopolized. Zhongtianwei is the only one in China. Self-Embedded CPU Intellectual Property (IP Core) Company. Data show that Alibaba invested in 2016, Zhongtianwei, SoC chips have exceeded 500 million in the past 2 years; Cambrian developed artificial intelligence chips for Huawei. Kirin 970 is famous in the industry.
Baidu has a relatively early AI R&D layout for graphics, voice, and driverless applications. In February last year, Baidu purchased 100% Raven Technologies. Last year at the Hot Chips Conference in California, Baidu released XPU, XPU is a 256-core, FPGA-based cloud acceleration chip, partner is Xilinx. In February of this year, Baidu acquired LightAlIence, an optical AI chip startup. Baidu’s AI strategy is to inject more technical power into the field of autopilot and smart voice. .
Tencent has always been in the hardware field has little movement Tencent has also been in the chip field has also moved, Tencent has invested in Bitland earlier, this company based on mining equipment requires a strong computing power has been developed based on ASIC TPU. Tencent and Alibaba They also invested in Barefootnetworks, which pioneered the new programmable PISA architecture. Its chips are mainly used in communications equipment.
In addition, according to statistics, Tencent has actually invested in Diffbot, iCarbonX, CloudMedx, Skymind, ScaledInference and other companies related to machine learning and artificial intelligence technology for several years ago, covering areas such as ubiquitous entertainment, health care, games, automotive transportation, etc. The game is based on Tencent's core products and social chain. It hopes to form a different business system and combine artificial intelligence to complete business innovation and breakthrough.
From a domestic point of view, perhaps the United States Department of Commerce’s export ban on ZTE has completely awakened the Chinese people and triggered collective anxiety about the core competitiveness of the Chinese chip industry. The chip investment theme has also begun to heat up and it has begun to be affected by the capital market. attention.
In fact, we thought that ZTE’s export ban should be more of a hardware company because of the ban on imports of components. However, the chip is always the most important one. But why did Internet companies start? Catch the chip this car?
First, improve computing efficiency and reduce costs
This may first of all be attributed to the changes in the chip industry, because AI chips are replacing traditional chips. The Internet giant is deeply laying out in the AI field. The so-called depth layout obviously needs to control the core components of both hardware and software. The software lies in the construction of the platform. The core of the hardware is the chip. To complete the chip ring, in order to have the opportunity to control the dominance of the future AI battlefield.
Raising computing efficiency and reducing costs may be the real reason for BAT to be a chip. Because AI chips can make artificial applications work better, they actually put these services on hundreds of computers in data centers. The cost is not cheap, but today's general-purpose chips have been difficult to meet the future demand for AI development.
In fact, as early as 2011, Google discovered that each user uses Google’s speech search service based on the deep learning speech recognition model provided by Google every day for 3 minutes. It must double the existing data center. This means that the existing CPU and The GPU can not meet the demand, which drives Google to develop more efficient chips.
In May 2016 Google I/O Conference, Google announced its own TPU for the first time. Google's senior hardware engineer Norman Jouppi said that Google TPU processing speed is 15-30 times faster than GPU and CPU (Compared with TPU is Intel HaswellCPU and NvidiaTeslaK80GPU), energy efficiency, TPU is an increase of 30 to 80 times. In the industry is relatively leading.
For Ali, the layout of the chip industry is closely related to its cloud computing business. According to the survey data of Synergy Research Group, in the fourth quarter of last year, Alibaba’s share of the cloud infrastructure services market was about 4%, compared to Amazon, Microsoft, IBM. There is a big gap between Google and Ali. Ali wants to bring new driving force to cloud computing through the layout of AI chips. At present, Ali's self-developed AI chips are mainly used to solve AI reasoning for image, video recognition, cloud computing and other commercial scenarios. The computational problem, in its official formulation, is a 'dedicated architecture chip' that is more dedicated, more optimized, has better performance, energy efficiency performance, and aims to achieve AI intelligence in business scenarios. In essence, it is still necessary to improve computing efficiency, reduce costs, and provide power for the data center.
If you want to play a tough game on the AI track, the chip is a key link - but the development of the AI depends on algorithms, data and power, and the chip carries the basic power, which is equivalent to the infrastructure, the chip. The pros and cons and the efficiency of operation are related to the power consumption, performance, and stability of the terminal side. No one wants to be stuck in this ring.
Second, combined with artificial intelligence, two-way fusion optimization based on algorithm and hardware design, gaining independent control
From a foreign point of view, it is worth noting that Microsoft and Amazon. In July last year, Microsoft announced at the CVPR conference held in Hawaii that they are developing new AI chips for HoloLens. The addition of AI chips will enhance their 3D gesture interaction and environment The deep understanding capabilities provide HoloLens with additional processing power, such as voice and image recognition, to support more tasks that gesture interaction cannot achieve.
Not to mention Amazon. As early as 2015, Amazon acquired an Israeli chip company called Annapurna Labs for $350 million. Annapurna Labs has developed a series of chips called Alpine that face memory, WIFI routers, and intelligence. Home, streaming media and other device types.
It is said that currently Amazon is developing AI chips for echoing Annapurna Labs. This chip will enable Alexa-based devices to respond faster and allow more information to be stored locally on the device without going through the cloud.
In addition, Apple built a 'neural engine' element for the chips in the iPhone X series mobile phones. A few days ago, Facebook is also forming an AI chip team.
Look again at domestic. In March 2017, Baidu released DuerOS smart chip, and reached strategic cooperation with Ziguang Zengrui, ARM, and Shanghai Hanfeng Electronic Technology Co., Ltd. This chip is equipped with DuerOS dialogue artificial intelligence operating system and will be given to the equipment. The ability to talk can be widely used in smart toys, Bluetooth speakers, smart home and other devices.
In Baidu's view, there have been two major artificial intelligence platforms, DuerOS and Apollo, which are also the inevitable choices under the combined strategy of hardware and software.
Tencent has invested in Bitland, which has released the intelligent video analysis server BF HS1, a deep learning server based on the BC SC1/SC1+ deep learning acceleration card and deep understanding of image recognition algorithms, specifically for video surveillance, Internet image processing. Various application scenarios provide deep learning acceleration.
Essentially, these giants are trying to solve the problem of deep learning, understanding, and having a faster response speed, whether it is abroad or domestic, so that artificial intelligence can tear off artificially retarded labels and become veritable.
For BAT, whether Baidu search, Street View, photos, voice interaction, driverless and translation, Tencent social networking, games, new retail, smart hardware, etc., or Alibaba New Retail, Cloud Computing, Internet of Things, etc. In the field, more and more algorithms and artificial intelligence technologies will be combined, requiring a deep learning framework.
Therefore, it can be said that the AI chip is driven by scenes, data, and algorithms. In the future, the amount of data processing at the application level will increase. Both require the optimization of bidirectional convergence based on algorithm and hardware design, and accelerate the speed of many artificial intelligence algorithms. Only by cutting into the AI chip can gain more independent control.
III. Anxiety under the Big Pattern of Sino-US Trade War: Decreasing Reliance on Peers
However, from the current situation, it is not optimistic. The main players in the global AI field track are concentrated in China and the United States. As mentioned before, in the industry chain, besides traditional chip giants such as Qualcomm Intel, the US Internet giant (Google) + Microsoft + Amazon + Apple + Facebook are all investing heavily in the chip field. Overall, the number and quality of companies in the United States at the algorithm, chip and data levels are upstream.
That is to say, current AI chips still cannot avoid the need for basic ICs such as CPUs and GPUs and existing operating systems to operate, while US companies have a very complete and balanced industrial structure in the areas of artificial intelligence chips CPU, GPU, FPGA, and ASIC, in GPUs and FPGAs. The field is almost monopolized.
This may be BAT's anxiety. Under the big pattern of Sino-US trade war, the United States will obviously encourage local companies to fully curb Chinese technology companies. For US-based technology companies, it is also their interest appeal. At the chip level It can make its hardware equipment more autonomous, get rid of or reduce dependence on chip companies such as Silicon Valley giants, and can have better differentiation from competing products, but this can't be done in the short term.
The current BAT strategy is biased towards wide net investment. Its purpose may be that once key chips are over-reliant on imports in certain key products in the future, it is time to look for domestic alternatives for verification, because maybe people will not know when they will be. It will be purchased at the national level. By the time you hold other suppliers in your hand, you will not be as good as ZTE.
After all, in the second half of the Internet or in the competition of globalization in the future, Amazon is targeting Ali. Facebook's competitor is Tencent. Baidu is facing Google.
Although hardware supply chain and software companies can achieve complementary business relationship chains and do not pose a core threat to each other's core business, if foreign Internet giants invest heavily, BAT certainly does not want to be subject to business in the chip business. Strongly related internet counterparts.
After all, at the AI track, no one wants to be the next ZTE. Then in the core chip business, even if it may not be as strong as the opponent, at least there are alternatives that can be on top, which is especially critical.
Third, non-mobile smart devices have opportunities to build new hardware and software platforms
From another point of view, although the software business often takes the form of gold, it is more profitable than hardware, but it has its own bottleneck and is subject to the mobile phone, the saturation of the PC hardware business itself and the disappearance of the demographic dividend. They all need Look for new growth points, and the new growth point may lie in the field of hardware and the field of things outside the smartphone.
Li Zhi, the founder and inquirer, once provided a set of data: At present, 95% of smart devices are smart phones, so voice interaction has not become the mainstream interaction mode. However, at present, the growth of non-mobile smart devices is far more intelligent than Mobile phones, and occupy more than 30% of smart devices in the next 3-5 years.
Where is the potential growth of non-mobile smart devices? A category that can be seen is smart speakers. Since there are approximately 40 million Americans in the United States who are users of smart speakers, equivalent to 20% of the total population of the United States, Compared with January 2017, the year-on-year increase reached 128%. In the smart speaker user group, Amazon echo accounted for nearly 70%.
According to Sun Yeh’s time machine theory, technological trends and trends in advanced countries will spread to less advanced countries in the next few years. Therefore, we have seen that BAT has now become smart speakers. At the beginning of March this year, BAT spontaneously took out its own Smart speakers and officially opened, Tencent Qrobot price 2799 in Jingdong; Baidu Raven H speakers can also be spot sales, priced at 1699; and Lynx mini version M1 has also been available. AI chip and smart speakers have what relationship?
Perhaps we can look at Amazon. We know that the mature AI chip layout can be applied to the voice interaction field of the Internet of Things or smart home. The smart speaker itself is also an important product to verify the scene and product of AI voice interaction technology. At present, Samsung, GE, Whirlpool, Lenovo, and other home appliance giants, household appliances covered by product categories (including televisions, sweeping robots, and smart audio) have launched products related to Amazon Alexa.
Shawn DuBravac, the chief economist of the Consumer Electronics Association (CTA), the organizer of the Consumer Electronics Show (CES) earlier this year, once said: 'Equipped with Amazon's Alexa Voice Assistant's products ... there are now about 1,500. With more and more related developers, the author pointed out that Alexa will gradually form a third-party open service platform in the future, just like the Apple App Store.
Therefore, this explains why BAT is not only doing AI chips, but also doing corresponding hardware. Because chips can provide computing power for hardware, and hardware provides grounding and control for their AI capabilities, and through Constant trial and error to verify the flaws of its data and algorithms and continuously improve its AI intelligence capabilities.
However, if simple data and algorithms cannot be combined with hardware to generate market explosions, their AI capabilities will hardly constitute absolute barriers and ecology, and it will be difficult to gain market recognition. Otherwise, they will be able to drive huge third-party vendors into their AI ecosystem. platform.
Through the convergence of AI chips and IoT devices, a new software and hardware ecosystem has been built out of the Android and iOS ecosystems. This may be the hidden aspiration of many software giants. After all, the case of overseas Amazon echo is just around the corner. Echo, as a hardware explosion model for its AI voice interaction capability, cannot drive the large-scale sales of its hardware products. It also introduced the entry of a large number of third-party vendors, and built a huge software and hardware based on Alexa. Ecological platform.
In general, AI and IoT began to collide and merge. IOT is the future trend. In the era of IoT, there will be new traffic entrances and even new platform-level entrances. The importance of AI chips is self-evident.
However, for BAT, if only the AI chip business is used as an investment project for trial and error, it does not mention the height of the strategic product to give sufficient attention to it. Therefore, it is difficult for the platform to invest in products produced by hardware manufacturers. Explosive models appear.
After all, from chip design, R&D to production, testing and verification to scale application is a very complicated system engineering. Especially, fully autonomous AI chips are highly dependent on cutting-edge talent, large investment, long technology research and development cycle, and many chip manufacturing technologies are also In the United States, it is not possible to pay for it. Although the self-controllability of chips is a deterministic medium-to-long-term trend, from the perspective of the self-controllability of high-end chips, it is still difficult for the listed companies and high-quality companies to take on this responsibility. Industrial chain clusters.
Compared with hardware vendors, BAT may have more chances in competition for talents. The innovation of AI chips involves artificial intelligence algorithms, programming languages, computer architecture, integrated circuit technology, and semiconductor technology. The key to making chips lies in talents. After all, the Internet giant's stable profits, good cash flow, and increasingly prominent brand influence have great advantages in the battle for chip talent.
Conclusion
The computer scientist Randall Allen K once said: 'People who really care about the software should do their own hardware'. This phrase is seldom noticed, but it has been verified by Apple and Amazon.
For the Internet giants, platforms that are truly integrated with hardware and software are often more sticky, have fewer short boards, and have more space for business divergence. They are the future growth point and they can also be better. Avoid being subject to humans.
Of course, the fact that software giants have made cores may have a direct impact on traditional chip makers. CNBC pointed out that you have chips, I have chips, and everyone has chips. This trend may eventually threaten big buyers and big supplies. The traditional relationship between merchants, especially Qualcomm and Intel caused a huge impact. But this does not affect the Internet giant's enthusiasm and dedication to the chip, of course, we are happy to see it.
Although in the context of economic globalization, no one company can completely rely on its upstream and downstream all-consuming, but at least it should be ensured that in the worst case, it will have its own cards and replacement options. This may be BAT to do The key reason for AI chips lies.