Since the beginning of this year, artificial intelligence has made breakthrough progress, and large models have shown various prosperity; In the industry, artificial intelligence has also begun to go deeper and serve more and more business scenarios.
On the one hand, artificial intelligence technology ushered in an inflection point, from perceptual understanding to content generation, from single mode to multi-mode, from dedicated to universal. On the other hand, the application of artificial intelligence has ushered in a singularity, and it is accelerating to enter the industrial production system. In the next two years, it will go deep into more than 50% of the core scenes of the industry, create trillions of dollars of new value, and gradually become the main engine of digital economic growth.
Industry intelligence is reshaping our times, and the intelligent future is accelerating.
Wang Tao, Managing Director of Huawei, Director of ICT Infrastructure Business Management Committee and President of Enterprise BG, delivered a keynote speech.
Industry intelligence faces four challenges: data, computing power, algorithm and application deployment.
However, the process of industry intelligence is not smooth sailing, and this process faces four challenges: data, computing power, algorithm and application deployment.
First of all, industry data is difficult to collect, transmit and use. Secondly, large-scale model training needs a lot of computing power, and enterprises need flexible computing power supply to meet diverse computing power needs. Thirdly, it is difficult for the basic large model to meet the accuracy requirements of enterprises, and a large number of industry large models and scene large models are needed. Finally, the business processes and business scenarios of enterprises are very complex, and how to deeply integrate AI with enterprise business is a problem that needs to be considered.
The key to speeding up the intelligentization of the industry is to make use of smart data, intelligent calculation and intelligent assignment.
In the face of the above challenges, the key to accelerating the intelligentization of the industry needs to start from three aspects: intelligent knowledge of everything, intelligent calculation of everything, and intelligent empowerment of everything.
First of all, Zhilian Everything, through intelligent perception and connection, realizes extensive collection and efficient transmission of industry data. Industry intelligence is inseparable from data, and the source of data is perception. In the past, the perception system was seven countries and eight systems, and the protocol was not unified, so the data could not be collected, collected incompletely and could not be read. In the era of intelligence, the perception system needs to be upgraded in an all-round way, so as to realize multidimensional ubiquity, open interconnection, intelligent interaction and easy-to-use intelligence.
For example, in order to realize open interconnection, HarmonyOS IOT operating system realizes data collaboration through distributed soft bus, unifies the language of sensing devices, and has the characteristics of super security, easy use and easy maintenance. HarmonyOS is moving towards various industry scenarios, and Huawei has developed "Mine Hong" for mining industry, "Dianhong" for power industry, and IOT operating systems for cities such as HarmonyOS.
Secondly, we can calculate thousands of numbers intelligently, and rely on a powerful computing base to tap the value of industry data, so that it can be varied.
The data should not only be collected completely and well, but also be connected widely and spread quickly, which requires high-quality connection to realize the intelligent connection of everything, ultra-wide flexibility, intelligence and self-control. At the same time, in the network convergence and backbone layer, the 400G/800G metropolitan area network enables enterprises to quickly connect clouds, calculate and combine intelligence, enabling "east counting and west counting", "east counting and west training" and "east counting and west saving".
Huawei’s high-quality 10 Gigabit campus network, with one network bearing the needs of office, production and Internet of Things. Extreme speed AP supports various access modes such as Wi-Fi, Bluetooth and RFID, and seamlessly connects various terminals. Intelligent SD-WAN gateway adopts optimized transport layer protocol, and the bandwidth utilization rate is increased by 8 times under the condition of 1% packet loss rate. At the same time, the end-to-end slicing network guarantees various business experiences; Network AI big model, improve predictive maintenance ability, nip in the bud, and ensure network uninterrupted and production uninterrupted.
Third, give wisdom to all industries, solve the problem that AI is difficult to land in industry scenes, realize the deep integration of various modes and industry scenes, and empower thousands of industries.
Together with its industrial partners, Huawei has jointly built an open and leading AI-based software and hardware platform, built a solid computing base in China and built a second choice for the world.
On the hardware level, based on Shengteng, Huawei and its partners have created a full range of artificial intelligence products such as modules, boards, training and push machines, servers, PoD and clusters, covering the whole scene of "cloud, edge and end". Artificial intelligence has entered the era of big model, and large computing power cluster is becoming the commanding height and core engine of artificial intelligence development.
The big computing cluster here is not a simple server stack, but a continuous breakthrough in key technologies such as computing architecture, ultra-high-speed network interconnection, large-capacity ultra-high-speed storage, and system-level ultra-high reliability.
Atlas 900 SuperCluster, a rising AI computing cluster with a new architecture, was launched.
As artificial intelligence enters the era of big model, big computing power is becoming the core engine of artificial intelligence development. Huawei changed the traditional server stacking mode, created an AI cluster with system architecture innovation, realized the integrated design of computing power, transportation capacity and storage capacity, and broke through the bottleneck of large computing power.
Facing the era of large model with trillions of parameters, Huawei launched ——Atlas 900 SuperCluster, a rising AI computing cluster with a new architecture, which can support large model training with trillions of parameters. According to insiders, it has three advantages.
First of all, it has achieved super-large scale.
Relying on Huawei’s long-term accumulation in the network field, the brand-new Huawei Xinghe AI intelligent computing switch CloudEngine XH16800 is adopted, and with its highest 800GE port capability in the industry, 18,000-card non-convergence cluster networking can be realized only by two-layer switching network.
Secondly, an innovative super-node architecture is adopted.
The cards are interconnected by 800GB high-speed bus to form a super node. Taking an 8-cabinet super node as an example, it can realize unified addressing of 544TB memory and unified scheduling of 200P FP16 computing resources. The whole cabinet adopts liquid cooling design, which realizes stronger computing power with smaller volume, and one cabinet tops two cabinets.
Thirdly, the system-level ultra-high reliability is realized.
Based on Huawei’s years of accumulation in computing, network, storage, energy and other fields, we will give full play to the comprehensive advantages at the system level, build an ultra-high and reliable training cluster, improve the reliability by 10 times, and improve the stability of large-scale model training from the sky level to the month level.
In addition, hardware enabling is the key to release large computing power, and it is also the basis of large-scale model development. In order to accelerate the innovation of big models, Huawei has released a more open and easy-to-use CANN7.0, which is not only fully compatible with the industry’s AI framework, acceleration library and mainstream big models, but also deeply open the underlying capabilities, so that the AI framework and acceleration library can directly call and manage computing resources, so that developers can customize high-performance operators and make big models have differentiated competitiveness.
Facing the development of Transformer network model, Huawei upgraded Ascend C programming language, simplified the logic of operator implementation with a more efficient programming method, shortened the development cycle of fusion operator from 2 person-months to 2 person-weeks, and enabled the rapid development of AI models and applications.
Work with customers and partners to release intelligent solutions for nine industries.
Aiming at the pain points of demand and doing practical things is not only the spirit of hard work that the whole society needs most, but also the purpose of Huawei’s digital and intelligent transformation of the industry.
In the past three years, Huawei has set up an industry corps to go deep into the industry and scenes, shorten the management chain vertically, and better respond to customers’ intelligent needs; Quickly integrate R&D resources horizontally and fully support the intelligent transformation of thousands of industries. At present, Huawei has built more than 200 intelligent solutions for more than 20 industries, such as cities, finance, transportation and manufacturing, and has been applied in a series of intelligent projects.
Based on the latest industry intelligent reference architecture, Huawei has jointly released intelligent solutions for nine industries, including finance, government affairs, manufacturing, electric power and railway. In the future, Huawei will continue to work with partners to create more industry solutions and promote the deep integration of AI and industry scenarios.
Launch the white paper "Accelerating Industry Intelligence"
Combined with the practice with a large number of customers and partners and the prospect of smart world 2030, Huawei also officially released the white paper "Accelerating Industry Intelligence".
The white paper proposes that artificial intelligence has become the core kinetic energy to promote social development and change, and is serving more and more industry scenarios and helping the industry upgrade. In order to better promote the innovative application of artificial intelligence, Industry-University-Research needs multi-party participation and collaboration, capacity co-construction and ecological co-creation, and under the premise of grasping the intelligent development trend of the industry, it will continue to pursue technological innovation, focus on engineering practice, and promote the AI application to go deep into reality, so that it can be varied and empowered.
The white paper has been recommended by many academicians, which embodies the experience and wisdom of all parties. It not only includes the development trend and latest trends of artificial intelligence, the application exploration of artificial intelligence in 63 scenarios in 16 industries, but also selects intelligent innovation practices in 18 industries, hoping to provide more specific help and reference for the implementation and landing of intelligent transformation in thousands of industries.
Intelligentization has opened a new chapter. Facing the opportunities and challenges of the intelligent age, let’s forge ahead hand in hand and build a solid computing base together, so that we can go deep into the scene in various ways, empower thousands of industries and accelerate the intelligence of the industry!
[This article is a speech by Wang Tao, managing director of Huawei, director of ICT infrastructure business management committee and president of enterprise BG, at the keynote of Huawei Connect 2023 on September 20th, 2023.]
Editor: Xu Yong