Shulou(Shulou.com)06/02 Report--

2019-10-12 20:41:26

Ren Zhengfei, president of Huawei, said in a speech entitled "how steel is made" in July 2019, "5G is a pediatrics and has been taken too seriously. 5G provides high bandwidth and low latency, supporting artificial intelligence, which is a big industry." As a traditional ICT business, Huawei's entry into the AI field is not hot-headed or blindly follow the trend, but out of in-depth consideration of its own business integration. As a technology outlet in today's science and technology world, AI not only has a broad potential market space, but also connects Huawei ICT, consumer electronics, IoT, cloud services and other businesses.

For the intelligent internal reference of this issue and my intelligent internal reference of this issue, we recommend the research report "Huawei's AI Vision" by Anxin Securities, which explains in detail Huawei's hardware and software AI system and reveals Huawei's AI strategic layout.

Hardware system: Huawei AI business foundation

1. Teng Teng: Huawei core focusing on AI scene

Teng 310-910, the bipolar system in the cloud has taken shape. On October 10, 2018, Huawei officially unveiled two AI chips, Teng Teng 310 and Li Teng 910, at the 2018 full connection Conference. Teng Teng will create five series of Max, Mini, Lite, Tiny and Nano to face different application scenarios. Among them, Teng Teng 310 belongs to Mini series, and Li Teng 910 belongs to Max series.

The optimal performance of all kinds of ▲ Teng chips across the whole scene

Teng Teng 310 is located in the edge side and end side AI chip, focusing on AI reasoning ability. The FP16 floating-point operation power of Teng Teng 310 reaches 8TOPS. Int 8 integer operation power reaches 16TOPS, the maximum power consumption is 8W, and the 12nm process is adopted. The Teng 910 is located in the cloud AI chip, focusing on AI training. The FP16 floating-point computing power of the Tengteng 910 reaches 256 TOPS and int 8 integer computing power reaches 512TOPS, and the maximum power consumption is 350W. The 7nm process is adopted. At present, the performance of the Teng 910 has to some extent exceeded that of the mainstream AI chips launched by Google and Nvidia.

Parameter characteristics of ▲ Teng 910 and 310

Horizontal comparison of ▲ mainstream AI chips

Da Vinci architecture, lock AI operator level optimization. For the AI computing scenario, Huawei has developed the Da Vinci instruction set architecture, which has the characteristics of high computing power, high energy efficiency and flexible tailoring. The main feature of Leonardo da Vinci's architecture is 3D Cube technology for matrix operations. Due to the demand of chain derivation of neural network in the process of AI computing, a large number of tensor (high-dimensional matrix) computation has become the bottleneck of the speed of traditional processors. 3D Cube technology accelerates matrix operations and enables the AI core to achieve 4096 MAC operations in one clock cycle, resulting in an order of magnitude improvement over CPU and GPU.

3D Cube Technology based on ▲ Leonardo da Vinci

CANN chip operator library, you must first sharpen its tools if you want to do a good job. CANN refers to the neural network computing architecture, which is the chip operator library and operator development kit located at the bottom of the chip. CANN has both optimal development efficiency and operator performance, and its core is the highly automated operator development tool Tensor Engine. Through the unified DSL interface, combined with pre-built high-level template encapsulation, automatic performance tuning and other tools, users can easily develop custom operators on the Teng chip. At the same time, CANN already supports all major AI frameworks. To develop this operator on the same chip, the general DSL needs 63 Loc, while Tensor Engine DSL only needs 22 Loc to achieve the same function, and the development efficiency is nearly 3 times higher. It is a powerful tool for developers to improve their efficiency.

Teng chip loading board, Atlas integrated products throughout the full stack application scenarios. Following the example of companies such as Google, Huawei does not regard Teng Teng chip as a separate product, but sells integrated products with integrated boards. At present, Huawei has launched a series of Atlas artificial intelligence computing platform products, with applications ranging from cloud side to end-to-side stack. At present, the traditional Atlas platform is mainly equipped with Teng Teng 310 chip, and its performance is limited. Huawei officially launched Atlas 900at Huawei full connection Conference in 2019. Atlas 900is the fastest AI training cluster in the world with several thousand Tengteng 910s. Atlas artificial intelligence computing platform products include:

(1) Atlas 900training cluster: targeted at large-scale problem AI training cluster server. The Atlas 900s are equipped with several thousand Tengteng 910s and take full advantage of Huawei's technology accumulation on the trunked communication library and job scheduling platform. The Atlas 900 has a total computing power of 256-1014FLOPS on 16-bit floating-point operations, equivalent to the computing power of 500000 PC. The training time for ResNet-50 model on ImageNet dataset is 10 seconds longer than that of competitors, which is currently the first in the world for AI cluster computing.

Time-consuming calculation of ▲ Huawei Atlas and similar products (seconds)

(2) Atlas 800 deep learning system: located in the cloud-side training and reasoning platform. Atlas 800 provides an optimized AI environment with pre-installed software environment for out-of-the-box use. At the same time, Atlas 800 integrates Huawei cluster management, task scheduling, performance monitoring and other functions, greatly reducing the threshold of enterprise AI applications. Atlas 800s are intended for AI developers and data scientists, providing integrated software and hardware delivery.

(3) Atlas 500 intelligent station: located on the edge side of the AI platform. Atlas 500 is suitable for transportation, nursing, unmanned retail, intelligent manufacturing and other application scenarios. It can achieve 16-way HD video processing capacity, which is 4 times higher than that of industry products. At the same time, Atlas 500 integrates two kinds of wireless data interfaces, WIFI and LTE, to provide flexible network access and data transmission schemes.

(4) Atlas 300AI accelerator card: it is a PCIe board, which can be used on the edge side server or on the cloud side through integrated loading. The Atlas 300 supports a variety of data precision and is equipped with a Penton 310 chip to achieve high performance computing, which is three times higher than that of similar products in the industry.

(5) Atlas 200AI acceleration module: located in end-to-side application scenarios. The Atlas 200 itself is extremely small, the size of half a credit card, and can be used for real-time analysis of high-definition video applications such as cameras, drones, robots and other applications. The power consumption of Atlas 200is very low, about 10W.

(6) Atlas 200DK AI developer Suite: targeted at end-to-side application scenarios and provided for developers. The Atlas 200 DKAI developer kit is built on the Penton 310 chip, and the core functions are open through the peripheral interface on the board, enabling "one-time development, multi-terminal deployment". The Atlas 200 DK AI developer kit is targeted at Ping an City, drones, robots, video servers, gates and other fields.

▲ Huawei Atlas series products

Deeply participate in the national project and establish close cooperation with Pengcheng Laboratory. Pengcheng Laboratory is positioned as the basic innovation platform to realize the national mission, focusing on the layout of the new generation of artificial intelligence basic research. Pengcheng Lab is currently focusing on building the Pengcheng Yunnao project. The basic platform it needs is the E-level AI supercomputing system. The built Yunnao 1 peak performance reaches 100PFLOPS, and Yunnao 2 plans to achieve E-level AI computing power. On the Pengcheng Yunnao supercomputing platform, Pengcheng Laboratory has created a Qizhi octopus computing engine, Kai Chi code hosting platform, digital retina and other products. Huawei Pengteng chip is deeply involved in the Pengcheng Yunnao project. At present, Yunnao 2 plans to adopt the Atlas 910 training cluster with Pengteng 910 as a whole, in order to achieve super computing power.

▲ Pengcheng Laboratory

▲ Pengcheng Yunnao

Teng Teng 610 is about to enter the smart car AI chip. At the same time as Huawei released the Penteng 910 chip, Xu Zhijun, vice chairman and rotating chairman of Huawei, also said that he also plans to launch the Penteng 610 in the future, mainly for self-driving scenes. In smart car products, Huawei unveiled the self-driving car brain MDC 600 as early as December 2018, and announced a high-profile cooperation with Siwei Tu Xin in 2019, which has been ordered by some domestic automakers. Under the blessing of the Teng chip, the performance of MDC will be further improved. At the same time, Huawei's Teng Teng chip development plan is being carried out in an orderly manner, and Teng 920 and Teng 320 will also be launched gradually after 2021.

2. Kunpeng: the national digital base of intelligent computing

At the full connection conference in 2019, Kunpeng chip once again became the focus, and a number of keynote speeches and keynote speeches were launched around Kunpeng. Kunpeng includes server and PC chips. Kunpeng 920 is the industry's first 64-core data center processor with 25% higher performance and 60% higher memory bandwidth than the industry's mainstream processors. At the same time, it combines CPU, bridge, network and disk controllers into one, making it the most integrated data center processor in the industry. Kunpeng chip develops according to the rhythm of "mass production generation, R & D generation and planning generation". It has been 12 years since 2007. Kunpeng 920 is now the third generation chip. We believe that Kunpeng chip is not only a digital base in the field of localization, but also a digital base in the field of intelligent computing, which is an important part of Huawei's strategy.

▲ Kunpeng 920

The evolution of ▲ Kun Peng and Gu Teng products

Perfectly equipped with Teng Teng chip, dual-engine drive intelligent computing. In large-scale data centers, the share of server costs continues to rise. According to Huawei cloud data center statistics, server costs have accounted for more than 60%. It is estimated that by 2025, AI computing power will account for more than 80% of the data center computing power. At the same time, the era of the Internet of everything is beginning, which will bring explosive growth of data, and massive storage and intensive computing will become a normal demand. IDC predicts that the future development direction of the computing industry must be the coexistence of multiple computing architectures, and the popularity of cloud services will accelerate this process. As a general-purpose computing chip developed by Huawei, Kunpeng can best match with the AI Teng chip to create a heterogeneous system of Huawei intelligent computing.

▲ Kunpeng + Teng Teng twin engines

Create integrated products to infiltrate both civil and commercial uses. Kunpeng is not only a single chip, but also has a lot of product expansion through integrated hardware. The main host of Kunpeng chip in the cloud is TaiShan series servers, which is currently the most powerful general-purpose server compatible with ARM architecture. For data center distributed evolution requirements and edge computing requirements. At the 2019 Huawei full connection Conference, Huawei once again exhibited the Kunpeng motherboard, including the PC side and the server side, and opened it to partners. Huawei Kunpeng motherboard adopts key technologies such as all-in-one SoC, xPU high-speed interconnection, 100GE high-speed iUniver and so on. It is not only equipped with Kunpeng processor, but also contains BMC chip and BIOS software. At present, Huawei has cooperated with partners such as Tsinghua Tongfang to develop its own brand of servers and desktop products.

▲ Kunpeng motherboard

Cost Analysis of ▲ data Center

Build Kunpeng computing industry, including high-quality upstream and downstream enterprises. OS based on openEuler and supporting database, middleware and other platform software are the foundation of Kunpeng computing industry. In order to expand the Kunpeng computing industry, Huawei has established an openeuler.org community, open source OS source code, and provided a variety of basic tools to shorten the development cycle for manufacturers to build OS, an openEuler-based distribution. At the same time, the Kunpeng computing industry will turn to support its partners, and the joint partners will complete the code migration and performance optimization of 3000+ applications in 9 months. Q1 plans to complete the construction of key industry benchmarking projects in 2020. 20% of the business scenarios can be migrated to the Kunpeng computing platform, forming a technical architecture deployment reference design for key businesses in key industries. Kunpeng has established extensive cooperation with a large number of A-share listed companies in server, database, public cloud and other fields to jointly promote industrial development.

▲ Kunpeng ecological partner

Issue a white paper and formulate a blueprint for development. Huawei released the White Paper on the Development of Kunpeng Computing Industry in September 2019, announcing its long-term development blueprint. An open and shared ecosystem is the basis for the success of Kunpeng computing industry. It is necessary to build the whole industry and Jingkunpeng computing industry system, and complete the goal of Kunpeng computing industry from the pilot of key industries to the construction of the industrial chain of the whole industry and the whole scene. Generally speaking, the plan is divided into three stages: pilot, promotion and deepening.

Pilot phase: through the selection of typical scenarios in government, telecommunications, finance and Internet industries for industrial enabling, incubation and pilot, through the pilot to build the confidence of upstream and downstream manufacturers and users in the industry.

Promotion stage: comprehensively open up the industrial system for government affairs, telecommunications, Internet, radio and television, financial securities, electric power, energy, transportation and other industries, providing a basis for digital business innovation in the industry.

Deepening stage: facing the whole industry, the whole scene, opening up the industrial chain and building an industrial system based on Kunpeng processor.

▲ Kunpeng development blueprint

3. Kirin: the pioneer in the battlefield of mobile phone AI

Many generations are iterated to achieve perfection. Kirin chip, independently developed by Huawei Hayes, has experienced more than 12 generations of product iterations. Kirin chip was the predecessor of Huawei mobile phone chip K3V2, which was launched in 2012. After two years of technological improvement, the first official Kirin chip product Kirin 910 was launched in early 2014. Kirin 910 is the world's first 4-core mobile phone processor, equipped with Mali-450MP4 's GPU. Since then, Kirin chip has launched a series of products for high-end applications and low-end application mobile phone products, which has greatly improved all aspects of performance. To Kirin 970MagneAI technology has become the standard Kirin chip, and continues to iterate to this day. The latest product of Kirin chip is Kirin 990. its first chip made by 7nm EUV process further enhances the function of AI.

Iterative Atlas of ▲ Kirin Chip products

NPU independent, to create an independent Kirin core. The AI function of Kirin chip was first supported by Cambrian technology. Through IP nuclear authorization, Kirin 970 and Kirin 980 are equipped with NPU designed with Cambrian 1A and Cambrian 1H architectures, respectively. Since Kirin 810 chip, Kirin chip began to carry NPU based on self-developed Leonardo da Vinci architecture. The latest Kirin 990 carries the NPU Macronucleus + NPU micronucleus architecture design, which is the branch version of the Teng architecture, Ascend Lite and Ascend Tiny. The transformation of Huawei in Kirin 810 and Kirin 990 chips fully shows that Huawei has got through the underlying technology of mobile phone AI chips, and Huawei's Teng architecture will gradually expand from cloud-side high-performance computing scenarios such as servers to end-side terminal devices.

▲ Kirin chip NPU architecture change

Kirin 9905G comprehensive AI ability first, highlighting the strength of AI. According to China Telecom's end-end Insight report in 2019 and AI evaluation tool AIT 3.0 independently developed by China Telecom, compared with other AI chip performance in the industry, Kirin 9905G performs the best in the industry in terms of backbone network model, multi-mode tuning, floating-point performance and hardware computing power. At the same time, Kirin 810 also shows strong AI capability.

▲ Kirin 9905G performs well

With the advent of MATE30, the three major AI functions lead the mobile AI application scenario. On September 19, 2019, Huawei launched the Mate30 series of mobile phones at the launch of new products in Munich, Germany. Among them, the front of the Mate30 Pro uses a "bang" screen design, which can be used for facial recognition through a 3D lens, while greatly adding sensor equipment. The three innovative features of Mate30 have become important highlights:

(1) Intelligent AI enables spaced operation. Traditional smartphones need a touch screen to operate the operating system, including clicking on the APP, taking pictures, screenshots and other functions all need the support of the touch screen. By contrast, MATE30 allows users to operate directly without touching the screen. For example, use both hands to make a fist to achieve the effect of screenshot.

▲ Mate30 AI spaced control

(2) AI enables intelligent tracking of binocular visual angle. During the use of traditional smartphones, the pictures or videos on the screen will change direction with the user's posture. Mate30 can intelligently track the eye angle and present it from the point of view that is most suitable for us to watch, thus freeing the hands of users.

(3) AI function protects privacy and security. Security and privacy is one of the key points that mobile phone terminal users pay attention to. Mate30's AI function allows the phone to automatically identify the reader when there is a new message, which makes users feel "sense of security" and "trust" to the phone.

Software system: the source of strength of Huawei AI

1. Create an independent AI framework MindSpore to open up the development path of the whole scene

At the same time as the release of the Teng series of AI chips, Huawei also released MindSpore, the AI development framework with Teng chips, making Huawei another technology giant with its own AI development framework after Google and Facebook. We believe that MindSpore has the following advantages:

(1) automatic and efficient parallelism to reduce the technical threshold of large-scale problem learning. Under the background of artificial intelligence accelerating landing, the scale of the data set is getting larger and larger, and the parameters of the model expand rapidly due to the increase of depth network layers. Traditional frameworks such as tensorflow and pyTorch use manual configuration to allocate computing resources such as CPU and GPU, which puts forward higher requirements for developers' understanding of parallelism and reduces the efficiency of development. MindSpore adopts the way of automatically allocating computing resources, which greatly reduces the technical threshold that users encounter when dealing with large-scale problems.

Performance advantages of ▲ MindSpore Framework

(2) optimize the architecture and design to improve the friendliness to developers. AI computing is highly dependent on chain derivation, so the automatic differentiation capability of AI development framework has become standard. Since the tensorflow framework was proposed, its graph-based reverse code scanning automatic differentiation has caused great inconvenience to users in the development process, so it has been widely criticized by developers. As a latecomer, MindSpore uses Source 2 Source to realize automatic differentiation, which is highly user-friendly.

(3) the combination of software and hardware to optimize the performance of the platform. One of the reasons for Google's strong AI cloud performance is that Google has greatly optimized the AI development framework tensorflow for self-developed TPU chips. Huawei, as a technology giant with self-developed chips, has also made a software and hardware optimization on MindSpore for Teng Teng chips to greatly improve the performance of MindSpore.

(4) implement logic improvement to fully protect the privacy of user data. With the explosive development of artificial intelligence, the issue of data privacy is becoming more and more important. Google, Apple and other big IT companies attach great importance to this issue, for example, Google launched tensorflow privacy on the basis of tensorflow. MindSpore has native support for privacy protection, by dealing with gradient and model information without privacy information, rather than the data itself, to protect the privacy of user data.

(5) to ensure independent safety and avoid potential risks. In our report "Open Source: free lunch?", we are the first to pay attention to the risks of open source projects in trade disputes, and further point out that AI development framework projects such as tensorflow and pyTorch maintained by foreign technology companies are dangerous high-risk areas. Therefore, Huawei's creation of MindSpore through self-research is the most powerful guarantee for its AI business in international trade disputes.

Horizontal comparison of mainstream AI development frameworks in ▲

2. AI development platform: 2 cloud-side platforms + 1 end-side platform

ModelArts: development platform + trading market. ModelArts, for developers, is an one-stop development platform for AI applications. The functions provided by ModelArts include massive data preprocessing, semi-automatic data tagging, large-scale distributed training, automatic model generation and so on. It has the ability to deploy end-edge-cloud model on demand, and can achieve full stack and full scene AI solution.

▲ ModelArts architecture

ModelArts has excellent computing performance and has topped the list twice on DAWNBenchmark. DAWNBenchmark is a standard computing power for image recognition application scenarios proposed by Stanford University, and it is also one of the main indicators for major manufacturers to show their AI platform capabilities. Up to now, ModelArts ranks first in the evaluation of DAWNBenchmark model training speed twice, reducing the time of using ResNet-50 to train ImageNet to 9 minutes and 22 seconds, and recently to 2 minutes and 43 seconds again.

ModelArts opens the AI model market to speed up the landing. AI Market is an ecological community of developers based on ModelArts, which provides content sharing functions such as AI models, API transactions, data, competition cases, and so on. Provide a secure and open sharing and trading environment for university scientific research institutions, AI application developers, solution integrators, enterprise and individual developers, effectively connect all participants in the AI development ecological chain, and accelerate the development and landing of AI products.

ModelArts is an AI development platform with automatic learning ability, which mainly has the following characteristics:

(1) automatic learning: ModelArts automatic learning ability, which can automatically carry out the whole process of model design, parameter tuning, model training, model compression and model deployment according to user tagged data. Without any experience in coding and model development, you can use ModelArts to build AI models for practical business applications.

Automatic learning ability of ▲ ModelArts

(2) AI full-process development: ModelArts full-process nursing can help you to complete AI development with high efficiency and precision, including data preparation, model building and model deployment. Data preparation includes three types of data types: image, sound and text, and seven types of annotations, which support the data requirements of mainstream models. The model building aspect supports an out-of-the-box online development environment, while simplifying and accelerating model development. In the aspect of model deployment, the model can be trained on ModelArts to support online reasoning service, batch reasoning task and edge device model optimization.

(3) Visual whole process management: full process visual management, no worries and focus on AI development. ModelArts provides visualization workflows from data, algorithms, training, models, and services. ModeArts provides a visual version comparison function to help users quickly understand the differences between different versions. After the model training is completed, ModelArts not only provides the visualization function of model evaluation in addition to the regular evaluation index display, but also provides the visualization function of model evaluation. You can understand your model vividly through confusion matrix and thermal map, and carry out evaluation model or model optimization.

HiLens: a visual AI development platform. Hilens is an AI development platform for developers specifically for visual application scenarios, providing one-stop AI application development, distribution, deployment and management, with massive device management, dynamic application deployment, online training, end-to-side model optimization and other end-cloud collaboration capabilities.

▲ HiLens platform architecture

Prefabricated skills simplify development, and the model is highly portable. Hilens provides a unified API interface, supports Caffe, tensorflow and other development frameworks, and simplifies the development process. At the same time, Hilens builds a developer community for developers to help developers share AI models. Hilens contains a wealth of AI skills, including human shape detection, face recognition and so on. HiLens supports the distribution of AI models to supporting hardware or any other device based on Huawei Hayes chip.

Equipped with supporting hardware, end-to-end cloud collaborative reasoning. Huawei provides Hilens with cameras with AI reasoning capabilities, HiLens discerning eyes. HiLens Vision is designed for developers to develop end-cloud collaborative vision applications, which is equipped with Huawei Teng chip and has a processing capacity of up to 100 frames per second. Hilens uses end-to-side supporting hardware equipment to cooperate with cloud-side algorithm model training, through end-cloud collaborative reasoning to balance low computing delay and high precision. Save the cost of cloud storage through end-to-side analysis of data.

Provide five application scenarios to accelerate the landing of visual AI. HiLens provides one-stop skills development. At present, it has provided five types of application scenarios, namely, AI skills development, campus scenario, family scenario, vehicle scenario and merchant scenario, which have basically covered the mainstream AI vision-related requirements.

HiAI: a lightweight platform for end-side AI. HiAI is Huawei's AI platform for Kirin chips and mobile phones. HiAI has completed two generations of product iterations. HiAI 1.0 was released in 2017 and is compatible with the P20 series of Huawei phones with Kirin 970. HiAI 2.0 was released in late 2018 to match Kirin 980 while adding a series of new features. At present, HiAI has made public 11 commercial landing projects, including SUNING, Kuaishou, WPS and other large enterprises.

HiAI is based on the "core, end, cloud" three-tier open architecture, from the bottom architecture to the upper architecture is HiAI Foundation,HiAI Engine and HiAI Service.

▲ HiAI three-tier open architecture

HiAI Foundation: the underlying architecture. HiAI Foundation opens the capabilities of NPU chips to users by closely combining with NPU on Kirin chips. Through rapid transformation, migration of existing models and heterogeneous scheduling, NPU achieves the best performance.

Highly adaptable Kirin chip to achieve automatic performance tuning. All the interface functions of HiAI Foundation have been integrated into Kirin chip, which can accelerate the calculation of neural network through HiAI heterogeneous computing platform, so as to realize operator interlayer fusion, efficient data placement, sparse model acceleration and so on. Users can focus on AI application development and leave performance optimization to the platform.

Have self-developed IDE to improve development efficiency. HiAI Foundation provides developers with a running environment and debugging tools based on mobile devices. At present, HiAI platform has integrated related functions into self-developed DevEco IDE. DevEco IDE is an integrated development environment provided by Huawei consumer business for developers to help developers quickly use the open capabilities of Huawei EMUI. By calling EMUI AI-related interfaces, mobile AI applications can be developed based on HiAI.

HiAI Engine: middle-tier architecture. HiAI Engine is built for the underlying software of each application scene. Through the construction of the engine library of three basic problems: computer vision, speech recognition and natural language processing, it realizes the open application ability and helps users to build fully connected services and panoramic applications.

Three major engine protection, the mainstream AI scenes are all covered. At present, the three basic scenarios of deep learning related applications are computer vision, speech recognition and natural language processing, which have been completely covered by CV engine, ASR engine and NLU engine provided by HiAI Engine. At the same time, based on the three major engines, HiAI Engine has derived seven categories of applications: face recognition, picture recognition, natural language processing, code recognition, video technology, speech recognition, text recognition, and 28 small categories of applications under it, covering all kinds of mainstream AI application scenarios.

▲ HiAI Engine function carding

HiAI Service: upper layer architecture. HiAI Service is essentially a platform with basic service capabilities. By opening the development interface to users, the service capability is open, and targeted push services are carried out according to the needs of users. HiAI Service is supported by two major platforms:

Fast service intelligence platform: Huawei's unified atomized service access and distribution platform acts as a traffic portal. Fast service smart platform provides a variety of access methods, including content interface fast service, card fast service, direct application fast service. At present, the express service intelligence platform already has successful cases such as Sina News World Cup, Baidu Map and so on. Fast service smart platform is not only a product of HiAI Service, but also an important support for Huawei's fast application strategy.

Xiaoyi Dialogue Development platform: Huawei's open voice dialogue service platform provides end-to-end development, testing and deployment of dialogue services for developers. Xiaoyi dialogue development platform has outstanding voice skills creation ability, can achieve convenient and smooth man-machine dialogue, and connect many voice skill creators. Xiaoyi dialogue development platform is not only a product of HiAI Service, but also an important platform for Huawei to grasp the voice entrance.

3. EI Enterprise Intelligent solution set: AI Agent + AI Special solution

EI Enterprise Intelligence Huawei Cloud provides the main bearer of semi-customized solutions for AI business. Huawei Cloud, with a revenue growth of more than 300% from 18 to 19 in China, has great potential for development and has received a large number of new orders from Internet, social consulting, video surveillance, car factories and other users. Huawei Cloud EI Enterprise Intelligence provides cluster services, and users have strong computing power. With the blessing of Atlas 900, it can achieve the fastest speed in the world, and can be used flexibly and on demand. EI Enterprise Intelligence launches related agents and special solutions for AI services.

EI enterprise agent is an industry universal solution built by Huawei based on Huawei Cloud and its own AI technology, with a high degree of freedom, and users can carry out customized transformation according to their own needs.

Traffic agent: located at using AI to create a safe, efficient and green urban traffic travel experience. The functions of traffic agent include global road network analysis, traffic prediction, traffic incident monitoring and control, signal light optimization, traffic parameter awareness and situation assessment, etc. At present, the transportation agent has been landed in Sakata, Shenzhen and Shangdi in Beijing.

▲ Huawei Cloud Industry Agent

Industrial agent: located in the use of AI to provide design, production, logistics, sales and service full-chain intelligent services, mining data value, helping enterprises to build a leading advantage with the help of new technology. The functions of industrial agent include energy saving and consumption reduction, process parameter optimization, predictive maintenance, sales prediction and so on. At present, industrial agents have constructed schemes such as optimization and improvement of product quality, maintenance of intelligent equipment, prediction of production materials and so on.

Park agent: located in the use of AI to deal with industrial park, residential park, business park management and monitoring, through video analysis, data mining and other technologies, to make production and life more convenient and efficient. The functions of the park agent include intrusion detection, lost property tracking, automatic face access control, intelligent parking guidance and so on. At present, the park agent has constructed the park access control, security area monitoring, intelligent parking and other schemes.

Automotive agent: positioned to support data access of on-board devices, provide real intelligent network connection, intelligent driving, UBI insurance and other scenario capabilities based on big data and artificial intelligence technology, to help the industry quickly upgrade intelligently and make people and cars live more intelligently. The functions of automobile agent include intelligent identification of safe behavior, real-time path planning, intelligent question and answer robot, real-time vehicle location insight and so on. At present, automobile agents have constructed biological keys, electronic fences, real-time path planning, intelligent dialogue robots and so on.

Network agent: aims at introducing AI into the network field, solving the problems of network business prediction, repeatability and complexity, improving the utilization rate of network resources, operation and maintenance efficiency, energy efficiency and business experience, and enabling the realization of self-driving network. The functions of network agent include data security, network experience embedding, rich application services and so on. At present, the network agent has constructed the schemes of data management, model development and training, network application service and so on.

▲ Huawei Cloud Agent

Third, why Huawei AI is "surging"

1. Vertical depth: the whole stack system of AI is vertically opened, with both hardware and software grasping.

Through the in-depth investigation of Huawei's AI product line, we believe that Huawei has built a full-stack AI product system, from the bottom chip to the upper scene application to achieve one-stop access. In terms of hardware, Teng Teng chip is the main line of AI product expansion. Huawei has not only built the Teng series of chips, but also successfully cooperated or carried them with chip products such as Kunpeng and Kirin, and further created integrated hardware products such as Taishan server, Atlas AI computing platform, Mate mobile phone terminal and MDC intelligent driving platform. In terms of software, Huawei Cloud is an important support for AI business. Huawei builds the MindSporeAI development framework according to the da Vinci architecture characteristics of Teng Teng chip, and builds three cloud and end-oriented AI platforms such as ModelArts, HiLens and HiAI. Relying on Huawei Cloud, Huawei has further expanded its EI enterprise smart solution covering transportation, medical, manufacturing and other application scenarios.

▲ Huawei Cloud Agent

2. Giant: learn from the development experience of major AI cloud service providers and innovate independently

Since the recognition accuracy of AI algorithm in ImageNet and other basic examples is higher than that of human dependence, AI as a major trend of the technology industry has gradually become the consensus of all giants. As a business with strong computing power, AI has a natural fit with cloud service products. As the importance of AI technology is gradually accepted by the public, the major cloud service technology giants have joined the bureau one after another. We have sorted out in detail the investment of the three foreign cloud service giants Google, Amazon and Microsoft in AI business:

Google: Google is one of the earliest players to enter the field of AI. Google's biggest advantage is algorithm accumulation and self-development of chips and frameworks. In 2016, Google's self-developed Alpha Go artificial intelligence platform successfully beat champion professional go players and published it as an academic achievement in the journal Nature, becoming one of the fuses to detonate fazhan in the AI industry. As a priority entrant, Google has a deep accumulation in the field of algorithms. at present, Google's latest automatic learning algorithm AutoML has become one of the leading AI technology trends in the industry. At the same time, Google independently developed AI chip TPU and AI development framework tensorflow. At present, tensorflow has become one of the most popular AI frameworks in the world, creating a strong ecological effect and cultivating the habits of a large number of developers. By combining hardware and software optimization, Google Cloud is one of the most competitive AI cloud services available.

Amazon: Amazon is one of the first companies to formulate the development path of AIoT. Amazon AWS's biggest advantages are IoT terminals and massive application scenarios. Amazon AWS launched AIoT products in early 2018, made a large number of optimizations for AWS, and integrated some pre-trained AI models into the platform for users in manufacturing, medical, security and other application scenarios. At the same time, Amazon launched the smart speaker Echo in 2014 in order to seize the AIoT entrance. Amazon's high share of the cloud services market and its priority layout in the AIoT field have established AWS's advantage in AI.

Microsoft: Microsoft's biggest advantage is the accumulation of algorithms in the AI field and its control of cutting-edge AR terminals. The Microsoft research team with Microsoft Research as the core has always followed up in AI English and published a number of articles in heavyweight scientific journals. Microsoft Azure has made some achievements in the field of voice and image. In terms of speech, it supports all kinds of speech recognition and natural language processing applications, and Microsoft Xiaoice conducts the diversion. In images, Azure can solve most computer vision-related problems and can interact with Microsoft AR device HoloLens.

Horizontal comparison of AI layout of major cloud service providers in ▲

By comparison, we can find that Huawei, as a laggard in the field of AI, has fully absorbed the relevant experience of foreign technology giants in the development of AI, and reflected the following advantages:

(1) full stack technology to build the most complete system in the industry. By comparison, it can be found that Google Cloud's leading advantage in AI business is not only due to its strong algorithm ability, but also closely related to its full-stack system. However, Google's full stack of technology products only stop at chips, frameworks, platforms, and do not constitute a complete closed loop. In contrast, Huawei has built a complete full-stack structure from instruction set architecture to industry solutions, thus gaining advantages in optimizing performance and reducing costs.

(2) to optimize the entry time and fully rely on the advantages of its own industry. As a traditional communication equipment provider, Huawei has a large market share in the field of cloud servers and mobile terminals. At the same time, Huawei has already launched HiLink and other Internet of things platforms to pave the way for the AIoT strategy. With the gradual maturity of AI algorithm, the innovation at the mechanism level of AI will gradually weaken, while the demand for application scenario expansion will gradually increase. This means that the opportunity to reverse infiltrate the AI market from devices, platforms and terminals has come, and the technology companies that can tamp the projects on the ground will stand out.

(3) to grasp the terminal advantage, there is the potential to grasp the entrance. Huawei Mate and Glory series mobile phones have always maintained a high market share and established a good user base. In the era of AIoT, mobile phone is a natural AI platform diversion tool. As a result, Huawei doesn't need to rely on other hardware products like Google does through Android, or cultivate users' habits through Echo like Amazon.

Potential entry advantages of ▲ Huawei Cloud AI Business

(4) the layout is far-sighted and closely follows the emerging areas. Since Huawei announced its entry into the AI field, Huawei has not been limited to traditional AI services such as recommendation system, security and voice recognition, but has always focused on the future. At the 2019 Shanghai International Auto Show, Huawei officially announced its entry into Intelligent Networked vehicles and made public its cooperation with Siwei Tuxin and other automotive networking advantages. At present, Huawei has launched the MDC intelligent driving computing platform, and will further improve the Teng chip to adapt to the smart car application scenario.

3. Forge a good relationship: establish an AI partner club and release the fertile soil Program 2.0.

Establish Huawei Cloud AI partner Club to promote AI business to become bigger and stronger in all aspects. In 2019, Huawei launched the global HCPN partnership system and launched the AI partner club program at the Hong Kong summit. AI partner Club Program is a global AI partnership program launched by Huawei Cloud to help partners build successful business or solutions based on Huawei Cloud EI products through training, technology, marketing and marketing. The specific projects of AI partner Club include AI Innovation Incubator, AI Industry Agent Alliance, AI Technical Cooperation Alliance and AI Consulting Service Alliance.

At present, dozens of innovative AI partners have announced that they have joined Huawei Cloud AI partner Club, covering healthcare, retail, campus, education, financial technology and other industries. Among them, NetEase youdao and Huawei Cloud jointly released NetEase youdao K12 smart education solution. At present, the disclosed members of Huawei Cloud AI partner Club include NetEase youdao, Jettong Huasheng, Yitko, Shenmu Information, Yi Tao Boshi, Anjie Zhihe, Shengmian Technology, Oneclick, Tableau Software, Hampen Technology.

Release fertile soil Project 2.0 to help AI grow ecologically. Huawei first unveiled a developer ecological strategy and a "fertile ground" plan for developers at the 2015 developer conference. Huawei will invest US $1 billion over five years to implement the fertile soil developer enabling program to create a development enabling platform and joint innovation for developer partners. At the technical level, Huawei provides developers with ICT open capabilities covering cloud computing, big data, the Internet of things and other areas. At the market level, Huawei focuses on business cooperation, supported by technical cooperation and personnel training, to help partners succeed in business.

After making some progress in the first phase of the fertile soil project, Huawei officially released fertile soil Project 2.0 at the 2019 full connection Conference, announcing that it would invest US $1.5 billion in the next five years to participate in the ecological development of the computing industry with individual developers and enterprises. Huawei will provide a series of resource support to developers, including Kunpeng development prototype, Teng training card, cloud service voucher, OpenLab, developer contest and so on. At the same time, three types of funds will be set up around the three stages of developer learning, product construction and product listing.

Zhi Dong believes that every IT giant has an "ecological dream". It is often said that "there are barriers only in depth", which can be understood as the depth of technology or product. The extreme of this kind of "depth" is "closed loop", or to build its own technology system. With the continuous expansion of Huawei's Penton chip and AI cloud services, Huawei's AI strategic framework has gradually become clear, that is to vertically break through the whole AI stack system and build its own AI closed loop.

Https://www.toutiao.com/i6746895778985280007/

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.