The next big trend and big bonuses from Internet plus in the AI +, has become the industry consensus. In the three swordsman of AI’s data, algorithms and chips, the data and chip will occupy a more and more important position considering the open source trend of the AI algorithm, and the chip of the AI development pillar is the competition of the AI industry.Contend for “core”. In the “movement” of a series of horse racing circles around AI chips, it is not a “single point battle” competition, but a full dimension of route, structure, application and ecology.
It can be said that the chip will determine the infrastructure and future ecology of the new AI computing era. Therefore, Google, Microsoft, IBM, Facebook and other American giants have invested heavily in the research and development of AI chips to seize the commanding point, while domestic AI chips are more innovative and more active.
For AI chip manufacturers, the primary task is to choose the right location. From the application point of view, AI chips are mainly divided into cloud AI chip and terminal AI chip. Cloud AI chips are applied to cloud servers and data centers; end AI chips are applied to intelligent devices and IoT terminal devices. In terms of function, training and reasoning need to be formedIt is.
In cloud training, multi GPU parallel architecture is a common infrastructure solution. In cloud recognition, based on the consideration of power and speed, heterogeneous computing (CPU+GPU+FPGA/ASIC) is the mainstream solution. And the terminal equipment focuses on inference. The person in charge of a IP company in the industry saidCloud training and reasoning are more performance oriented, without considering power consumption, memory and other factors. In the terminal training, it requires a lot of data and a lot of computation. For AI chip companies, future opportunities are definitely terminal reasoning.
Manufacturers anticipate different future plans. Some need to make big chips in the cloud. Some are facing terminals, developing specific scenes, and others are concurrent. What kind of route do you choose? Who will win the “bets” of GPU, FPGA and ASIC? Whether it is based on GPU, FPGA and other general-purposeThe semi custom solution of the chip is still a dedicated ASIC chip, and the final competition is still cost-effective.
Dispute of Architecture
On the AI chip, there are both Google, Intel, NVIDIA and other leading giants, and countless small and medium enterprises are going all out. The struggle behind the line is the struggle between the traditional and the new architecture.
ICCADThe director and Professor Wei Shaojun of Tsinghua University said that structural innovation is an unavoidable problem. At present, the AI scheme architecture used by each family is incompatible with the standard AI computing interface that is not supported. Will there be a universal processor like CPU in the same time to the unified AI chip world? If it exists, architectureWhat is it? If it does not exist, the AI chip that currently meets a specific application must only exist in the way of IP core, and eventually integrated with a wide variety of SoC. If so, what will the company that is engaged in AI chip development today?
AIThe basic elements of the chip are: programmable, dynamic variability of architecture, high computing efficiency, low cost, simple application development, and so on. The current CPU+GPU, CPU+FPGA, CPU+ASIC and so on are not ideal architectures. Professor Wei Shaojun believes that the general AI processor drivesThe power comes from the following two aspects: the cost of developing the AI chip hardware architecture is very expensive, not all the manufacturers can bear; the scalability of IP, the support breadth of the architecture, the standard AI computing interface are very important to the popularization of the AI chip, and the development related algorithms can not be completed on the eve of the day.
As one of the main players on the AI chip circuit, Cambrian believes that AI will inevitably have a heavyweight application, and the hardware must be standardized and universal. The founder of the horizon, Yu Kai, thinks that the application of AI is vastly different in various scenarios, and more likely, there will be one in every subdivision.A leading structure. The future architecture debate will continue.
Battle of camps
Under the huge draught of AI, veteran recruits were flocking in to share the dividends of the new ecosystem, and also magnified the variables again.
In addition to the emerging horizon, Cambrian, deep seal, bit continent and other manufacturers, we see that, on the one hand, many old chip enterprises are actively embracing AI, becoming an important force in the AI chip industry, such as the traditional SoC processor chips, such as Zhongxing, Beijing Jun, Zhongtian micro, Hangzhou wick, etc.The multimedia chip enterprise is the representative. On the other hand, the algorithms and systems companies that cut into the AI chip are increasing, such as Shang Tang, Hai Kang, Dahua, Yi, HUAWEI and so on. They are actively investing in AI chips through independent R & D and M & A investment.
Compared with the active new AI chips, the old companies have a more complete team of front and back design, product, verification and test, and have the engineering experience of building a complete SoC chip product. The entry competition of the old chip enterprises and the wrestling of the new strength in the talent and product level are the future.The AI industry is a major concern. In the near future, system companies have a deep understanding of the real needs of the scene, strong integration of hardware and software, marketing channels and sufficient capital reserves, which will make them always in a relationship with a large number of AI chip start-ups.The more dominant position has increased the uncertainty of the industrial structure.
AIThe technical barriers to chips are not low, but as long as the industry concentration is high, winners will choose to eat all. For example, the manufacturers of mobile phones, shipments to a threshold, have the power to make their own chips, such as apple, Samsung, HUAWEI and millet have chosen to develop their own mobile chip. This is for Qualcomm, the coalition, the display, and so onThe machine chip supplier is also a big impact. And this is going to be the same story in the AI field.
The dispute of application
Commercial application is one of the key factors of AI. AI has value only if it solves practical problems. The huge R & D cost, chip cost and distribution cost of AI chips need to be used to “dilute”.
But according to expert opinion, the key application of AI needs 99.9. There are more than 9 of the percent, and they can’t be commercialized if they can’t do it. For example, the common feature of autopilot, “key applications”, is that the project is usually expensive, the R & D cycle is long, the technology bull is needed and the financing can be sustained.Power。 Most of them are non critical applications, such as face recognition, usually combined strength, including insight, product and engineering capabilities, cost control, supply chain capability, marketing ability, iterative ability and so on.
In addition, different industries have different “focus”. In the automotive industry, safety and real-time are the most important issues. In the field of security, the most important thing in AI+ video surveillance is to turn passive monitoring into active analysis and early warning, which requires high recognition rate, calculation power and cost. In the field of consumer electronics, mobile phones areThe HUAWEI mate10 with Unicorn 970 chips and iPhoneX, which also embed AI chips, lead the cell phone into the intelligent era. In addition, Amazon’s Echo detonated the smart home market, and the demand for AI is to solve the problems of power consumption, security and privacy.
Which vertical field to choose depends on some key factors: is the market space big enough? What is the degree of concentration in the industry? Is technology an improvement or a revolution? Who is higher on the competitors’ barriers? Obviously, there are still many tough battles to be applied to AI scene applications such as consumer electronics, security and smart cars.
The dispute of ecology
On the competitive dimension of AI, ecology is definitely a key link.
Under the trend of AI platform, the future AI will present a number of dominant platforms and the competitive pattern of extensive scene application, and ecological builders will become one of the most important models. At present, technology giants are already laying down the layout of the basic technology and application level of the AI industry chain, for example, Google has launched Ten.SorFlow distributed learning framework, the domestic Baidu established and opened the PaddlePaddle open source deep learning platform, and also introduced the DuerOS and Apollo two AI operating systems.
For Tencent, Alibaba and Baidu giants, in order to maintain their status as a king, AI ecosystem must also be constructed. At the ecological level of AI, what is worth paying attention to is that, with the evolution of the ecology, many fields will have a trained model to refer to, and the barriers to future algorithms will be used.More and more low, if the core competitiveness of a company is only an algorithm, it will be very dangerous; two is that AI chip vendors can not only consider their own unique architecture and software development suites, and in the current situation, they will not be able to build full AI service flow capabilities, including hardware and software ecology, or need to be integrated.A suitable ecological circle, otherwise the long-term competitiveness will be difficult to guarantee.
With the evolution of AI, the improvement of computing power has promoted the development of algorithm. The development of algorithm has raised the demand for AI chip. “You are running, and the original advantage is running.” This sentence applies to the battle between giants of infield and Intel, and also to AI chip companies. In the new AI chip domain, the future editionThere are still many imaginary space to see how to ink. (proofreading / Fan Rong)