Computing: The Engine of Intelligent Twins

Computing is at the heart of empowering organizations to deploy intelligence.

By Deng Taihua, President, Huawei Computing Product Line
Deng Taihua

By leveraging diversified computing power (general-purpose computing + heterogeneous computing), the platform allows collaborative innovation across the full stack, from hardware and basic software to the application enablement. 

As a general-purpose technology, AI is drastically changing the way we live and work and accelerating the deployment of intelligence in industries. Intelligent upgrades are crucial to the continued competitiveness of every city, industry, and company. Intelligent Twins form an integrated intelligent system comprising cloud, edge, and networks, serving as the optimal route and a reference architecture for governments and enterprises to become intelligent.

With computing power as the engine of Intelligent Twins, Huawei is committed to working with its partners to develop Kunpeng-based computing (general-purpose computing) and Ascend-based computing (AI computing) to seize the strategic opportunities of the new era. Diversified computing power (general-purpose computing + heterogeneous computing) will facilitate collaborative innovation across the full stack, from hardware and basic software to application enablement, catapulting governments and enterprises into an intelligent world.

Ascend-powered AI 

The Ascend-based computing platform exclusively provides basic software and hardware to China's Ministry of Science and Technology's National Open Innovation Platform for Next Generation Artificial Intelligence.

The platform boasts Atlas series hardware, the Computing Architecture for Neural Networks (CANN), MindSpore AI computing framework for all scenarios, and MindX for Ascend application enablement. On the platform, AI developers, researchers, and industry application partners can more effectively develop scenario-specific AI solutions.

The Atlas series of hardware covers all scenarios across the edge and cloud to support AI inference and training, and offers a broad portfolio including modules, cards, AI servers, and clusters. Following its strategy of opening up hardware and enabling partners, Huawei cultivates certified solution partners and integrates Ascend modules, acceleration cards, and servers into its partners' own products and solutions. So far, Huawei has developed five solution partners.

Communicate 87

CANN – a heterogeneous computing architecture – is crucial to unlocking hardware computing power. Its southbound APIs are compatible with various hardware form factors on devices, the cloud, and the edge, and support 14 major operating systems. In addition, its northbound APIs support the industry's mainstream AI frameworks. With the unified programming interface AscendCL (ACL), partners can use a full range of AI inference and training hardware for simplified development. The performance of Ascend hardware is maximized through coordinated software and hardware optimization, Ascend-friendly diagram compilers, and over 1,000 high-performance operators. These advantages allow CANN to help our partners develop and provide unrivaled performance.

MindSpore is an AI framework designed for device, cloud, and edge scenarios and serves as the key to ecosystem development. MindSpore features automatic parallelization, end-to-end simplicity, and adaptive, all-scenario collaboration. With growing AI models, the number of GPT-3 parameters has exceeded 175 billion, requiring future frameworks to provide higher parallel computing capabilities.

As the industry's first automatic parallel framework, MindSpore can divide large diagrams into smaller ones, automatically split operators within these smaller diagrams, and then assign the operators to different computing units. All of this is possible without the intervention of developers. After models are trained, MindSpore generates different-sized models by adapting to different hardware form factors, and then automatically adjusts these models.

MindSpore also supports training of local incremental data and provides precise AI models that can meet developers' requirement for customized frameworks. This allows developers to rapidly import and debug models, and use tools to automatically switch between models that are trained by third-party frameworks.

Since going open source this March, MindSpore has received high acclaim from global developers. The Mindspore community now has over 1,500 open source contributors, and daily downloads in September exceeded 600. It’s becoming the most popular AI open-source community in China and is seeing a growing number of downloads outside of China. Huawei looks forward to working with AI developers to build a global mainstream AI framework that can adapt to different scenarios.

MindX features two components for deep learning and intelligent edge, one select model library, and software development kits (SDKs) for various industries. The development toolkit and industry knowledge library for one industry can be replicated in other industries for easier and more efficient development, with developers able to deploy AI functions with few or even no lines of code using MindX SDK.

Open collaboration 

As part of its strategy to open up its software and hardware and enable partners, Huawei has made the full-stack AI software and hardware platform open source for the entire industry, aiming at building an Ascend-based AI industry ecosystem. To enable industry applications and empower industries with AI, Huawei has partnered with over 100 leading global independent software vendors (ISVs) and developed more than 120 solutions using Ascend. Through the concerted efforts of over 100,000 developers, AI applications jointly developed by Huawei and its partners have been widely adopted by the transportation, financial, energy, manufacturing, Internet, and telecommunication industries, among others, and have delivered tremendous value.

The Manufacturing Dept is key to Huawei's core competitiveness, and is responsible for manufacturing Huawei's ICT infrastructure and smart devices. Huawei ships hundreds of billions of units every year, and produces one phone every 28 seconds and one piece of base station equipment every three minutes. Despite being a leader in both China and worldwide, quality control of its numerous products is still a crucial task. Many production processes used to be inspected manually and re-inspections were often required to ensure accuracy.

To solve this problem, Huawei's Manufacturing Dept worked with ADLINK to integrate AI into quality control and launched the Atlas solution with cloud-edge synergy. This solution was first piloted on the production line for computing products – the production line saw a 60 percent reduction in inspection workloads, with accuracy jumping to almost 100 percent since the pilot began in June 2020. Following this, Huawei began applying the solution to over 80 additional production lines, including wireless and devices. The Atlas solution helps Huawei extend beyond traditional manufacturing to intelligent manufacturing.

Based on its manufacturing know-how and AI development experience, Huawei released intelligent manufacturing SDKs and opened them up to its partners to enable efficient intelligent manufacturing and AI application development. Huawei's Atlas solution lowers the barriers to developing AI industry applications, and allows smart factories to flourish.

Communicate 87

A single architecture cannot support the diversified computing requirements of diverse scenarios like big data and AI. Huawei aims to enable the intelligent upgrade of governments and enterprises through diversified computing power. Based upon China's New Generation Artificial Intelligence Innovation and Development Pilot Zone, Huawei has built computing power clusters like AI computing centers, and established diversified public computing power platforms to bring together enterprise innovations and empower the intelligent upgrade of industries. On the edge, Huawei has developed edge AI computing devices, such as Atlas 200 modules and Atlas 500 AI edge stations, and launched over 100 industry AI solutions with its partners. These solutions include intelligent grid inspections, intelligent service centers, and intelligent gas stations. Computing power clusters will empower industry clusters and boost regional industries thanks to cloud-edge-device synergy between the clusters and the edge.

Developing talent as the cornerstone of the intelligent world

Talent development is the foundation of industry development. As part of its work to cultivate the computing industry, Huawei has placed talent development high on its agenda. Universities are a major source of talent, and the future of talent development. Huawei has worked with universities to establish a knowledge system that adapts basic theories to industry and market needs, so as to develop talent by combining theory, practice, and application.

Huawei has partnered with experts and scholars from various domains and Tsinghua University Press to publish a series of textbooks and education materials regarding Ascend AI processors, MindSpore, ModelArts, Kunpeng processors, openEuler, and openGauss.

Supported by China's Ministry of Education, Huawei has also signed agreements with more than 70 universities to build an industry-university talent development base and promote Kunpeng and Ascend in its universities. So far, Kunpeng and Ascend have become courses at more than 2,600 Chinese universities, colleges, and vocational schools, and are gradually being adopted outside China. Moving forward, skills relating to Kunpeng and Ascend will become essential for AI, computer, and information majors.

An open and shared ecosystem is critical to Intelligent Twins. Looking ahead, Huawei will continue investing in the R&D of basic computing software and hardware, and opening the full-stack Ascend and Kunpeng platforms to the industry to foster a diversified computing ecosystem. This will provide the computing power necessary to build new infrastructure and move towards a high-quality digital economy.