Cloud computing is integrating with the continuous development of artificial intelligence technology , It is accelerating the efficiency of enterprise development .
3 month 27 Japan , Baidu intelligent cloud “ 2021 Cloud Intelligence Technology Forum ” The first event was held in Beijing , Baidu Intelligent Cloud heavy release “ Cloud intelligence integration AI Developing full stack mode ”, Based on unique AI Native Cloud Architecture , High performance through integration of hardware and software AI Developing infrastructure to provide services for enterprises AI Develop full stack solution , And shared with the guests of enterprises in various fields AI The characteristics and practice of Zhongtai, etc . As a domestic AI The Baidu cloud is the first in the public cloud service market “ Integration of cloud and intelligence ” Exclusive advantage continues to lead the market .
* Poke video to see the wonderful scene of cloud Intelligence Technology Forum *
Baidu intelligent cloud is Baidu AI Carrier and exporter of technology landing , Has shown “ Integration of cloud and intelligence ” This unique competitive advantage .“ wisdom ” Baidu brain , As the core foundation of Intelligent Cloud , After more than ten years of accumulation , Baidu brain has developed into a core technology leading software and hardware integration AI Mass production platform , Empowering for all walks of life AI New infrastructure .
In order to meet the large-scale production of industrial intelligent application , Baidu intelligent cloud is the first to put forward in the industry. AI Native cloud computing architecture , Create the most suitable running environment in the industry AI Cloud , Combined with the upper layer, it is flexible and easy to use 、 To meet the application needs of various developers AI Development platform , Help enterprises to improve AI Development efficiency , Speed up the intelligent process .
The present , Smart home 、 Smart finance 、 Autopilot 、 Intelligent medical treatment 、 Smart retail 、 Industry applications in intelligent manufacturing and other fields are entering AI The primary stage . In order to better adapt to Should have come AI A new stage of primordial development , Baidu Intelligent Cloud building full stack AI Native Cloud Architecture , oriented AI In depth software and hardware collaborative optimization of application scenarios , Build the industry's most suitable for running AI Cloud , stay AI The infrastructure level can provide enterprises with high-performance AI Calculate the force , more Simple 、 Easy to use 、 Efficient development process , Smoother end-to-end application development whole process experience .
Baidu intelligent cloud is the main link too AI Calculation 、AI Storage 、AI Containers Three layer construction AI Development of infrastructure . The specific term ：
► AI Computing layer
Based on Baidu “ Taihang ” Elastic bare metal products provide customers with high performance , Multi specification and cost-effective computing services . This is divided into the following dimensions ：
Chip level , Rich in AI Chip type , Including Baidu's self-developed Kunlun chip and multi specification commercial products in the industry GPU and FPGA chip , To meet the needs of different scenarios AI Computing power demand .
Single server layer , be based on X-MAN Architecturally AI Superserver , X-MAN It's the first one to match 4 road CPU Super AI Computer , breakthrough CPU+GPU Matching limit , Better release GPU Ability , Give Way CPU The ability to be GPU Constraints on the utilization of computing power , promote AI Overall computing power of computing scenarios . At the same time, communication between multiple cards in a single machine is based on NVLink High speed interconnection .
Baidu AI Super server unified through Baidu independent research and development of hardware virtualization technology Baidu Taihang elastic bare metal , Take Baidu Taihang high performance computing as an example , Seamless access to other cloud services , Including cloud storage 、 Cloud database 、 Cloud network, etc , At the same time, it also keeps the high performance of zero loss at the bare metal level .
► AI Storage layer
Its core is based on Baidu's object storage service and related object-oriented service AI Scene acceleration and processing capability , Solve customer problems AI Data storage and analysis of development scenarios . So , Baidu Intelligent Cloud provides special orientation. AI Computing based storage service architecture , And divided into the following aspects ：
first floor , It's data on the cloud , How to help customers' data go to the cloud quickly and efficiently , Baidu not only provides disk transmission mode for offline transmission , Online oriented cloud flow product , Support customers to create online transmission tasks of various models . Offline transmission supports one hundred TB Data migration capabilities , Online data flow cloud flow Compatible industry 7 Object storage interface .
The second floor , It's data storage , Baidu takes object storage as the core to provide cost-effective storage services . Baidu object storage provides the industry's first 4 Level data storage , And provide intelligent life cycle management and industry-leading read-write performance , Meet the cost performance requirements of customers in the big data scenario .
The third level , It's data processing and processing AI Acceleration layer , At the acceleration level , Baidu provides one layer Cache service , Storage media based on higher speed , Provide metadata acceleration , Cache management and other capabilities , To provide higher performance storage capacity , At the same time, some of them are integrated on the object storage by default AI Intelligent processing capability . Baidu's cache capability , In some scenarios, training speed can be improved 4 More than times , At the same time, it also integrates dozens of intelligent processing capabilities .
Besides , For video 、 Image and other file types , Baidu Intelligent Cloud in BOS Test better package integrated Baidu related AI Ability , Including dozens of image auditing capabilities , And image enhancement and special effects capabilities , Users can go directly through BOS Event triggered framework for , During data access or data upload , Call these intelligent processing capabilities directly through unified interface expansion , Data analysis and storage closer , Reduce coding complexity , Improve management experience .
► AI Container layer
Baidu Intelligent Cloud's container engine service CCE, Providing the basis for Docker Container lifecycle management 、 Operation and maintenance management of large scale container cluster 、 Business applications can be released and run with one click , Also for AI The scene has been optimized , Include GPU Management of heterogeneous resources ,AI Job scheduling management, etc , To use computing resources more efficiently , Improve the utilization rate of resources . Tell me more about AI Some key features of Container Services ：
Support GPU Sharing ability of , Support users to carry out 1/2、1/4 Equigranular GPU Computing resource management .
stay AI Scheduling level of job , Support Gang、Spread、Binpack And so on , And aiming at NVLink etc. GPU Aware scheduling capability of network architecture .
At the acceleration engine level , Acceleration and communication engine , Operator acceleration can improve reasoning efficiency several times in specific scenarios , The communication library can support the communication acceleration capability of kilocalorie scale .
With the tide of industrial intelligence ,AI It has become a common technology for the transformation and upgrading of many industries , Enterprise class AI The role of development platform is increasingly prominent . Depending on the AI Native cloud infrastructure , Based on the industry level open source deep learning platform developed by Baidu Flying propeller , Meet the customized application requirements of enterprises for scenarios , Baidu launched Flying propeller Enterprise Edition , Including facing AI Zero threshold created by application developers AI Development platform EasyDL And face AI Full function of Algorithm Developer construction AI Development platform BML.
EasyDL and BML Can satisfy the data acquisition of developers 、 feature extraction 、 Parameter tuning 、 model training 、 Model to evaluate 、 Model deployment 、 Reasoning service and other whole process requirements , More abundant task scenarios 、 Convenient and efficient intelligent data service EasyData、 Built in Baidu super large scale pre training model brings high precision effect 、 Four core advantages of model deployment supporting extensive adaptation to various types of hardware .
EasyDL oriented AI Application Developer , Support image 、 Text 、 video 、 voice 、OCR、 Structured data 、 Retail industry 7 Big technology direction ,16 Types of tasks , Support for public clouds 、 Local service deployment 、 Device end 、 Four deployment schemes of software and hardware integration , Has served more than 80 Million users . from EasyData Intelligent data services 、 From model training to service deployment, one-stop whole process service allows users to understand the details of the algorithm ,5 Minutes to go , The fastest 10 Minutes to complete model training .
► Data preparation
On data processing , The intelligent data service platform can be used directly EasyData The power of , Realize data collection 、 assessment 、 cleaning 、 One stop service marked . Greatly reduce the cost of data acquisition and processing .
► model training
EasyDL Built in the super large-scale visual pre training model developed by Baidu and the pre training model Wen Xin of natural language processing （ERNIE）2.0, Compared with the pre training model of open source dataset training, it can effectively improve the effect of the model .
EasyDL It also released a new multi angle model evaluation capability to further help model tuning , It includes three visual attribution analysis tools ：
Confusion matrix ： Through data statistics , Help developers accurately determine which categories of models are prone to errors .
Heat map ： It can further explain the decision basis of the model , The importance of pixels that affect the result of model recognition is given in the whole picture .
Noise sample mining ： According to the confusion matrix and thermodynamic diagram , Directly mining the noise samples that affect the effect of the model , Help developers to carry out model iteration in a targeted way .
► Model deployment
In order to meet the needs of different scenarios , Baidu provides public cloud API、 Local server deployment 、 Device end SDK、 Four deployment modes of integrated hardware and software products . Users only need to go through simple settings , The trained model can be transformed into a service that meets the needs of the business scenario . At device side SDK On , Adapted NV Jetson series 、Intel Nerve accelerator 、 Huawei NPU、 Huawei Atlas、 qualcomm DSP、RK Such as more than 15 Four mainstream chips and four operating systems , Realize the widest adaptation in the industry .
in addition , EasyDL Recently, the end cloud collaborative deployment scheme was launched ：
End cloud collaborative deployment is based on the intelligent edge of Baidu open source BIE Frame building , Developers can easily manage all kinds of terminals and edges in the cloud .
On end and edge devices , Integrate edge suite once , In the local , You can easily get AI Model local reasoning 、 Video stream access and other functions . And in the clouds , The service can be distributed and updated , There are also report statistics and node management functions .
Collaborative deployment of end cloud is very suitable for those with networking （ Or partial networking ） Business scenario under the condition of business environment , Developers train models , as long as 1 minute , New models can be distributed and updated to the edge for verification or batch deployment . It can be greatly improved AI Deployment efficiency of model development , So that developers can focus on the business itself .
EasyDL The deployment scheme of end cloud collaboration can run on all kinds of intelligent terminals , Like a box 、 Integrated machine 、 In the server , On these edge devices , All kinds of sensors can be connected again （ Like a camera 、 MIC）.
combination EasyDL Public cloud deployment and local server deployment , It is convenient to build a complete set of cloud edge fusion AI Solution , Meet the requirements of various business scenarios AI Application requirements .
BML It's for AI Algorithm Developer building , In model construction and model management ,BML Provides cost-effective computing power 、 It is easier to use various modeling methods . Model storage is also provided for model management 、 Model transformation 、 Function of model evaluation and model optimization . so to speak BML For professional algorithm developers, it provides more flexible and powerful AI Development platform .
BML It has the following four core advantages , Comprehensive modeling methods 、 Preset high performance AI Kit 、 Flexible delivery 、 Provide a variety of domestic solutions , To provide enterprises with independent, controllable and widely adaptable AI Development platform .
BML Provide preset model parameters 、Notebook、 Custom assignments 、 Visual modeling and other modeling methods . Its in , Preset model tuning is based on low code modeling , Have some flexibility in script editing , At the same time, the built-in automatic super parameter search function helps to improve the effect of the model .BML The automatic super parameter search function provided is an innovative parameter adjustment algorithm without gradient optimization based on stochastic differential equation , Fast convergence , Independent of smoothness assumption , And it can support large-scale parallel search parameter adjustment . When you turn on auto super After the search ,BML The average accuracy of online multi scene model can be improved 10% above .
► High performance AI Kit
High performance machine learning Suite . Acceleration in machine learning scenarios for data analysis , A suite can produce ten times the effect of an open source Suite , Its interface is very compatible , With a high-performance machine learning Suite , In several common open source tasks, the performance has been greatly improved .
Flying propeller Literary heart ERNIE Development Kit . Flying propeller Semantic learning model for Sustainable Learning , In addition, it also provides 30+ Algorithmic tools ,20+ Refined pre training model ,20+ Tools and methods of data processing 20+ Tools for basic development , The algorithm that makes the whole data annotation input 、 The development time has been greatly reduced , The efficiency has been greatly improved .
BML There are four delivery methods to meet different needs , Including the public cloud 、 Private cloud 、 A hybrid cloud 、 Integrated machine . The all-in-one machine provides a general rack type server scheme and a whole cabinet scheme based on Scorpio , For overall power consumption 、 performance 、 Heat dissipation has been optimized , And encrypted by model 、 Code reinforcement 、 TPM Trusted authentication 、 Host protection and other ways to strengthen the security .
► Localization solution
BML Full support from domestic deep learning framework Flying propeller , To Kirin and other domestic operating systems , Then to China CPU and GPU And the Great Wall 、 dawn 、 lenovo 、 Introduction of various hardware forms , It forms a self-control system 、 Wide range of applications BML Integrated machine , Cost effective computing resources meet all kinds of computing needs .
Enterprises are in the process of intelligent upgrading , Especially when the application develops from a single point vertical scenario to a comprehensive upgrade phase of multiple business scenarios , Enterprises will also face challenges from AI Capacity production 、 There are many problems in application, operation and maintenance and management , At this time, enterprises will need AI Zhongtai , As a set of enterprise's own AI A platform for capacity production and centralized management .
AI Zhongtai is an intelligent center of enterprises , Connecting data within the enterprise 、 Knowledge and business , Faster and more efficient support for the upper level of intelligent business applications and business innovation . Baidu intelligent cloud AI The core of Zhongtai includes AI Capability engine 、AI Development platform , Combined with the supporting data management 、 Service management 、 The whole line resource operation and maintenance management system, etc , Enterprises can build a set of intelligent upgraded infrastructure . stay AI Capability engine , Enterprises can start from Baidu's existing 270 Multiple maturity AI Application of direct selection in ability .
AI The development platform includes EasyDL、BML And scene based customization platform UNIT etc. , Meet the multi-level development needs of enterprises . Baidu Intelligent Cloud combined with specific industry characteristics , Energy has been created 、 Finance 、 City 、 Media and other industries AI The solution of middle platform landing , Provide construction services for enterprise customers in different industries AI Independent ability of development and Application , Intensive management of enterprises AI Capacity and resources , The effective path of overall planning enterprise intelligent upgrade layout .
When AI and traditional industries accelerate convergence , Enter fast 、 The stage of large scale application , Baidu based on itself AI The leading potential energy formed by the long-term accumulation of technology and ecology will be released , To meet the real needs of the intelligent era “ Integration of cloud and intelligence ” Enabling industry in a new way .“ Integration of cloud and intelligence ” Become the core advantage of Baidu Intelligent Cloud , It's also a hug AI The best choice for enterprises and developers .