Cover Story
The Smart Future Is Here: Winning the Next Decade on HUAWEI CLOUD AI


On October 10, 2018 at the HUAWEI CONNECT summit, Huawei launched its Full-Stack, All-Scenario AI Solution as Huawei's formal introduction to the AI scene. Huawei's Rotating CEO Eric Xu described Huawei's business strategies in the development and adoption of AI technology.

Huawei's AI strategy is based on continuous investment in cutting-edge research and finding the newest talent to help develop the newest technologies. The Full-Stack, All-Scenario AI Solution utilizes state of the art resources to provide a reliable, groundbreaking solution, helping to develop a global ecosystem and allowing customers to harness the power of AI. 

Full-stack is designed to improve efficiency vertically throughout the profile, and comprises chips, their enablement, and training and inference framework. All-scenario refers to horizontal frameworks, meaning this solution is perfect to improve deployment environments such as public cloud, private cloud, edge computing, IoT devices, and so on.

The Full-Stack All-Scenario AI solution is powered on HUAWEI CLOUD. Although the cloud service brand was established just a couple of years prior, it inherits the most complete tech portfolio in the industry forged from 30+ years of accumulated expertise in serving carrier, enterprise, and consumer customers by its parent company. Providing key services in software, hardware, and solutions, has caused Huawei to become a key industry pillar, and HUAWEI CLOUD BU one of its star players.

The general manager of EI service products at HUAWEI CLOUD, Jia Yongli said, “From day one HUAWEI CLOUD has brought the advantages of full-stack to everyone”. With full-stack technology at its core, this past year alone HUAWEI CLOUD served 8 major industries in over 200 enterprise projects software and hardware integration and for all infrastructure layers. It combines industry knowledge with cutting-edge AI technologies to help such industries as transport, logistics, and manufacturing gain the competitive advantages they need to stay at the top of their game.

All-Around Improvements in AI Capabilities on HUAWEI CLOUD

Based on the new Ascend chip, HUAWEI CLOUD upgrades the performance of the infrastructure at all layers, from the bottom layer to software framework to networks, creating a solid foundation on which to run AI services.

First tier evolution: bottom-layer compute enhancements

Computing power, algorithms, and data are the three cornerstones of AI. Currently enterprises produce vast amounts of data from cleaning, marking, and training, all of which cause a long delay. In most cases, computing resources are scarce for such tasks.

The foundation of computing power is chip. At the HUAWEI CONNECT 2018 summit, Huawei introduced two AI chips, the Ascend 910 and 310, which created quite a stir. These two chips leverage Huawei's unique “Da Vinci” architecture.

Ascend 910 belongs to the Max series and features ultra-high computing power for cloud scenarios. It is the world's largest AI chip to run on 7 nm technology, with its half-precision computing capability reaches 256 TFLOPS (tera floating-point operations per second), double that of the 125 TFLOPS rating of NVIDIA V100, and has a maximum power consumption of 350 W. The Ascend 910 chip will be launched in the second quarter.

Ascend 310 is a Mini series chip that uses 12 nm technology and provides efficient computing and low power consumption. It features 16 TFLOPS computing power, a maximum power consumption of only 8 W, and an integrated 16-channel HD video decoder. The system on a chip (SoC) is the industry's most powerful SoC in edge computing. 

It is widely known that Chinese cloud vendors do not produce their own quality chip, instead buying from third-parties. Some vendors are in development, but it is still way too early for such chips to enter service. Huawei is already using its Ascend 310 chips in its delivery of cloud services, becoming the standard configuration in its high-end VMs, benefiting users with 16 times the inference power.

High-Performance Computing H6 ECSs now come standard with 16 TFLOPS AI inference capabilities. For those requiring even higher levels of technological firepower, HUAWEI CLOUD also provides 512 TFLOPS computing services in AI-enhanced VMs and containers. In training scenarios, HUAWEI CLOUD provides bare-metal versions of VMs. A single node can provide a maximum of 2 PFLOPS of computing capability. 

In addition to the H6 service, HUAWEI CLOUD also provides Ai1, At1, and Physical.At1 services, all of which are powered on the new Ascend AI chip technology. With the groundbreaking series of cutting-edge cloud services leveraging these chips, HUAWEI CLOUD has made its strong presence known, leaving vendors in the race to catch up. 

Jia Yongli noted, "Computing power has become a significant business focus for Huawei. It is definitely not the only competitive advantage of HUAWEI CLOUD for artificial intelligence, but rather one of countless advantages."

Second tier evolution: full platform training and inference framework

On top of the compute/operator advantages, Huawei also released a complete software stack to yield the unique end-to-end optimizations that come from a truly full-stack portfolio. 

Most AI computing algorithms need to be trained on the cloud end before being deployed at the terminals. However, many vendors do not provide chips for terminals, and likewise chip vendors do not provide such cloud computing services. Most AI application data needs to be migrated between training and deployment. For enterprises, this puts them in a precarious position, but more often than not, it is a waste of time, money, and effort. 

The conversion of underlying layers covers many algorithms and operators, meaning that despite its initial appearances that the system is running smoothly, the efficiency is low. Huawei hopes to make this easier by implementing a single framework covering public cloud, private cloud, edge computing, and remote devices that can run in a variety of AI environments. It ensures AI can be set once and run without further configuration, making life easier for the developers. 

CANN and MindSpore constitute the core basic framework of the Full-Stack All-Scenario solution.

The operator library CANN provides optimal development and operator performance, including the Tensor Engine, which uses uniform DSL interfaces, as well as automatic operator tuning, generation, and optimization. In the Tensor Engine, Huawei uses the TVM concept proposed by Chen Tianqi and others. Take the Reduce sum development case as an example, using CANN improves the development efficiency by three times.

MindSpore is an AI framework proposed by Huawei for unified training and inference that supports deep, reinforced, and enhanced learning. It can flexibly adapt to different resource budget deployment environments, provide consistent development experience at the device, edge, and cloud ends, and support all mainstream machine learning and deep learning frameworks (including TensorFlow, PyTorch, PaddlePaddle, Keras, ONNX, and MXNet). It is due to be released in the second quarter of 2020, at which time Huawei will also introduce the device's deep learning framework. The size of the framework is only 2 MB, and the memory usage is less than 50 MB when MindSpore is running.  

Together, CANN and MindSpore form the core framework for the Full-Stack All-Scenario solution.

Tier three evolution: one-stop AI application development platform

In addition to the scarce and expensive computing power, the AI industry's large-scale development is hindered by low development efficiency. The entire process of marking, training, and deploying an application is both time-consuming and labor-intensive. To solve this problem, Huawei launched the ModelArts platform.

ModelArts is a one-stop development platform for AI developers. It provides data preprocessing and semi-automatic marking for massive volumes of data, large-scale distributed training, automatic model generation, model optimization, and on-demand deployment of device-e-cloud models.

In addition, ModelArts can provide visualized management of the entire lifecycle of AI development, covering original data, annotation data, training jobs, algorithms, models, and reasoning services. It manages tens of millions of models, data sets, and services, generates source tracing diagrams without manual intervention, and selects any model to find the corresponding data sets, parameters, and models. In particular, the training comparison result function is highly acclaimed, even among Huawei internal developers.

ModelArts Yields Four Main Areas of Acceleration

Fast data preparation

Data annotation and preparation are the most painful tasks for AI developers. 70% of the overall development time using conventional methods was occupied by data preparation. ModelArts has a built-in AI data framework, which uses a groundbreaking AI mechanism and iterative training platform to manage data and solve the problem of data volume. For scenarios with a large amount of data, ModelArts improves data annotation and preparation efficiency 100-fold.

Fast learning

ModelArts provides an automatic learning function, supports automatic model design and parameter adjustment, and allows developers to quickly roll-out their applications.

Another key component is MoXing SDK, which supports rich model libraries, optimization algorithms, and various tool libraries, allowing developers to compile and import algorithm code. ModelArts supports automatic parameter optimization as well as training, verification, prediction, and model export. Developers only need to write a set of code and ModelArts can automatically implement its roll-out to single, multiple, and distributed nodes.

Perfect even for AI beginners who want to quickly generate models, ModelArts also provides algorithm models such as RestNet_50, Faster_RCNN, and SegNet_VGG_16 that cover most common application scenarios. All preconfigured models are trained using open source data sets and have a leading model precision. Training algorithms can be implemented with just one-click, where you only need to configure the data path, log output path, and Hyper Parameter automatic selection to start the training. In the future, more algorithm models will be launched to suit a variety of demanding scenarios. 

Fast training

To tackle the challenge of time-consuming model training, ModelArts uses the same model, data set, and equivalent hardware resources with its state of the art optimization technologies, including cascading hybrid parallel technologies, to reduce the model training time by half.

For many carriers, large-scale cluster deployment where large amounts of data needs to be synchronized is a big issue. Currently, the best achievements of the distributed training of big data sets have been that of the team operating on Amazon's cloud service, who completed training within 18 minutes using 128 GPU instances simultaneously. However, lab results show HUAWEI CLOUD EI can use the same type of node to complete the same tasks within 12 minutes.

Fast rollout

For scenarios where AI is largely adopted, model deployment is very complicated, time-consuming, and labor-intensive. For example, in the Smart Transportation field, updated models need to be rolled out to cameras of different specifications and vendors simultaneously.

Using this example, ModelArts can push models and updates to all edge and end devices with just one-click. Cloud deployment also supports online and batch inference to meet requirements of multiple scenarios, such as large concurrency and distributed scenarios.

Huawei HiLens, a Development Tool Purpose Built for Visual Intelligence

HiLens is composed of an AI capability camera and a cloud development platform that provides powerful computing and storage space. It can implement 100-frame processing capability and millisecond-level face detection, meeting the requirements of large image processing.

In addition, HiLens is a lightweight container that features low resource usage, low network bandwidth, and fast download, which reduces the difficulty of real-time system processing.

HiLens provides a complete set of secure and reliable one-stop skill development, deployment, and management services for individual developers, enterprises, and device manufacturers. It can seamlessly connect to customers' industry devices and improve development efficiency and productivity.

HiLens provides rich models and skills for control codes and models. ModelArts can first train any AI models, which then is implemented as the basic development component for AI functions, which in turn are then directly deployed on any device using the Ascend chip. HiLens is compatible with other mainstream framework training models. When deployed on the device side, these third-party models are automatically converted into the MindSpore model to provide optimal performance and interoperability. 

For conventional AI implementations in visual intelligence use cases, different processing methods are advantageous for research and development than other scenarios. However, problems occur in complex and demanding application environments. For example, the recognition rate of the face recognition system for machine vision detection can reach as high as 95% in ideal environments, but when tested in the field, the recognition rate of visual detection is greatly reduced. HiLens provides online training device-specific models online based on the unique deployment data of each device, improving model precision and enhancing user experience.

ModelArts and HiLens went live on HUAWEI CLOUD EI platform on the opening day of the HUAWEI CONNECT 2018 summit.

In addition to the groundbreaking developer-oriented tools listed above, HUAWEI CLOUD EI's entire bucket is much more comprehensive than it was last year.

Currently, HUAWEI CLOUD EI has launched 142 functions from 45 services. For solutions, HUAWEI CLOUD EI provides three types of services: General API, high-level API, and pre-integrated. Customers from wide-ranging technical backgrounds can find something suitable to them and their demanding service environments. Data scientists, data algorithm engineers, IT developers, and even business personnel without AI background can find easy-to-use solutions covering most cloud scenarios.

HUAWEI CLOUD aims to build an all-around AI ecosystem that streamlines underlying hardware and upper-layer software applications, developers, and industries. Further it does not focus solely on ideas, strategies, or slogans. This is the foundation of what HUAWEI CLOUD has done and continues to do.

Intelligent Twins for Urban Management Applications in Huawei EI Suite

The Huawei AI Full-Stack All-Scenario Solution was launched at HUAWEI CONNECT 2018 and is generally available for commercial use. ModelArts, HiLens, and other developers from the full-stack dimension were also exhibited, attracting large interest from developers. Based on Huawei Ascend series chips and mainstream heterogeneous computing components, the Atlas intelligent computing platform has implemented AI capabilities in cloud, edge, and device through various product forms, such as modules, boards, small cell, and all-in-one machines.

At the same time, HUAWEI CLOUD EI industry solutions were released and on display. These solutions cover many industries such as public utilities, transportation, finance, logistics, education, and retail. HUAWEI CLOUD EI has successfully completed projects with major carriers, including Microsoft and Intel, and many enterprises specializing in industry intelligence have demonstrated the intelligent transformation to their services thanks to HUAWEI CLOUD EI.

In September last year, HUAWEI CLOUD launched their Enterprise Intelligence (EI) services, including basic platform, general services (big data, visual cognition, and voice semantics), and industry-specific solutions. Later in 2018, HUAWEI CLOUD EI Enterprise quickly launched its Smart Water, Smart Manufacturing, Smart Electricity, Smart Transport, Smart Finance, and Smart Retail solutions, making them available for commercial use.

By integrating industry know-how with AI, HUAWEI CLOUD AI acts as the catalyst for industrial upgrades in three main scenarios: repetitive and high-volume work, high-value work, and multi-domain collaboration. By specializing in these domains, HUAWEI CLOUD AI can break through barriers to help improve efficiency across industries, transfer expertise from experts to common users, and break the limits of human intelligence.

For repetitive or high-volume workloads, HUAWEI CLOUD EI services help identify massive sets of frequently-used data in enterprise practices. For example, two HUAWEI CLOUD EI services – DLS and Image Search – have helped TukuChina automatically import and cross-check hundreds of thousands of copyrighted images and tens of millions of images from the Internet every day, with an accuracy of 99%. 

For high-value workloads, AI that is loaded with expert experience or industry insights can act as assistants to experts. The HUAWEI CLOUD EI team has been working closely with KingMed Diagnostics, a company that provides medical diagnostic testing. By leveraging AI technologies, KingMed Diagnostics made breakthroughs in the pathological examination of cervical cancer, improving the reliability of in-house diagnosis with a sensitivity (true positive rate) of over 99% and a specificity (true negative rate) of over 80%.

For tasks that require multi-domain collaboration and involve many parameters, complex dependencies, and high dimensions, such as industrial production and urban governance, AI brings with it new ideas and methods. HUAWEI CLOUD AI works as a passageway to a new world of groundbreaking technologies that are, for most carriers and small-scale businesses, unattainable due to the prices. 

The HUAWEI CLOUD EI City Intelligent Twins project is a typical example of the multi-domain collaboration scenario. Based on digital twins, AI collaborates with multiple technologies, such as cloud, big data, edge computing, and Internet of Things (IoT), to achieve a complete closed-loop system covering data generation and analysis. Through the powerful computing power of the digital world, the physical world is more intelligent.

At present, City Intelligent Twins continuously explores efficient resource scheduling and configuration in transportation, emergency services, environmental protection, water conservation, and energy to solve more urban problems. When put into practice, this solution improves efficiency, reduces city-wide energy consumption, and improves environmental protection.

Li Qiang, the director of the traffic police department of Shenzhen Municipal Public Security Bureau, shared several figures of real data that demonstrates the significant changes brought by HUAWEI CLOUD EI City Intelligent Twins.

Since its adoption in the first half of 2018, HUAWEI CLOUD EI City Intelligent Twins helped Shenzhen traffic police deploy AI applications that monitor traffic violations, such as using mobile phones when driving and not wearing safety belts, as well ensure that those accused successfully face law enforcement. The traffic police law enforcement volume in Shenzhen increased by 15%.

Shenzhen traffic police are deploying the EI traffic intelligent TrafficGo solution at 43 intersections in the Bantian and Longgang districts in Shenzhen. This solution has shortened the average waiting time of cars at key intersections by 17.7% when the online signal configuration is piloted.

Since the construction of a new operation command center, the emergency response time of Shenzhen Traffic Police is shortened by 67%.

Another case of HUAWEI CLOUD AI technologies being deployed in demanding situations is that of Shenzhen Airport. This airport has a record over 1,000 flights a day, foot traffic of over 120,000 passengers, and a direct boarding rate from terminal (bridge rate) of approximately 70%. Comprehensive technology was needed to keep up with the demands, and +AI is used to implement intelligent reconstruction of the airport's legacy infrastructure. Using Gantt charts and AI automation, the bridge rate is increased to 80%, an increase of 10%, which means that there will be 4 million people who do not need to take a bus to get to their seat. At the same time, Shenzhen Airport is working with facial recognition to achieve one-stop customs clearance, aiming to reduce the queuing time of passengers by 15%.


In the future, migration to the cloud will become the norm. Because AI requires massive computing resources and storage space, cloud will be the only environment for most enterprises to explore artificial intelligence. "Cloud + AI" is the driving force behind industry transformation.

CTO of the HUAWEI CLOUD BU, Zhang Yuxin, believes that the advent of the Cloud 2.0 era is significantly different from the Internet development of the past decade.

First, he points the trend of how enterprises are actively moving to the cloud, especially when migrating key applications. Whereas in the past, cloud has been used only in personal entertainment and consumption fields, in the Cloud 2.0 era, cloud has entered the production field.

Second, he notes how the frequency and severity of bottlenecks that occur in traditional Internet applications are no longer acceptable for large enterprises. 

Previously, internet applications mainly relied on traffic dividends, meaning that business growth correlated to the amount of traffic. However, the traffic dividend will eventually have a ceiling, and the traffic dividend model is easy to copy, something other businesses are likely to adopt. In the Cloud 2.0 era, the core of new Internet services has caused a movement from traffic dividends to data dividends.

Zhang Yuxin believes that in the Cloud 1.0 era, technical keywords, such as distributed, automated, and large-scale resilience, were used to resolve customer problems, meaning that problems were categorized into one of these three pockets and customers were relatively naïve to any alternatives. However, this is not the case in the new era because such troubleshooting is far from enough. In addition to security and reliability, enterprise applications and Internet applications require intelligent intelligence, big data, and technology and system architecture. Cloud 2.0 is the time for AI, for intelligent industries and carriers to find value in their vast volumes of data. 

At HUAWEI CONNECT 2018, many technologies and their applications were demonstrated. The O&M platform and tool framework of the chip, hardware, data center, basic cloud service, and application development platform were developed and put on display for the whole world to see. Similarly, innovative basic AI model algorithms together with industry-oriented industry solutions quickly became the talk of the industry, and will no doubt continue as Huawei keeps it foot on the accelerator. HUAWEI CLOUD has made many technical breakthroughs to cope with the challenges of enterprise intelligence of today, and tomorrow.

In the Cloud 2.0 era, many surprises await us. HUAWEI CLOUD is here to guide you through the wilderness.

Provides Scalable, On-demand Computing Resources

Learn More