Reshaping Industries with AI: Huawei Cloud Tackles Big Challenges Head-on
Sep 22, 2023 GMT+08:00
[Shanghai, China, September 21, 2023] On September 20, HUAWEI CONNECT 2023 kicked off with the theme of "Accelerate Intelligence". Joined by thought leaders, business elites, technical experts, partners, and developers, the event dives into the latest advances in business, industry, and ecosystem.
September 21 was off to a good start with the keynote speech by Mr. Zhang Ping'an, Executive Director of Huawei and CEO of Huawei Cloud. In "The Cloud Foundation for an Intelligent World: Reshaping Industries with AI", he announced Huawei's Pangu models for mining, government, vehicles, weather, medicine, virtual humans, and R&D, as well as the brand-new Huawei Cloud Ascend AI Cloud Service. These products and solutions demonstrate Huawei Cloud's efforts in promoting cloud-powered AI across industries.
"We tackle big challenges head-on." Mr. Zhang said, adding that Pangu focuses on the most pressing needs different customers face and reimagines work and life for every industry, every enterprise, and every individual.
Zhang Ping'an, Executive Director of Huawei and CEO of Huawei Cloud
Reshaping Industries with Pangu Models
Mr. Zhang first highlighted the achievements of Huawei Cloud in boldly facing the past year of obstacles.
This February, Huawei Cloud launched 23 software development services to cover code repository, requirement management, and testing. Partnering with more than 20 top industry players, Huawei Cloud released cloud native hardware development tools for schematic diagrams, PCB layouts, and structure design. This June, Huawei Cloud announced GaussDB for redefining database building, a distributed cloud database already widely used in Huawei and key industries such as banking, insurance, securities, and energy. In addition, Huawei Cloud MetaStudio, a one-stop digital content production pipeline, provides as many as 10,000 cores to render one image. With this amount of computing power providing a fundamental boost in digital content production, 3D movie rendering now takes just one week down from six months.
Looking back the HDC.Cloud in this July, Huawei Cloud released Pangu Models 3.0 to speed up AI industry adoption. Its 5+N+X decoupling architecture and layered AI capabilities support a vast range of models and applications. HUAWEI CONNECT is where Huawei Cloud further extends Pangu models into mining, government, vehicles, weather, medicine, virtual human, and R&D.
The first major industry application is in mining, an industry pressured to boost cleaned coal production. Shandong Energy Group adopted the Pangu Mine Model in nine processes and 21 scenarios, including digging, drivage, equipment control, transport, ventilation, and coal washing. They separate coal using the prediction model. The Group's No. 2 coal mine in Jining now produces an additional 8,000 tons of cleaned coal each year, equivalent to 2‰ more yield and CNY20 million more revenue. Pangu Mine Model currently works in eight mines in China.
In a second industry application, Pangu also streamlines government services. City governance requires collaboration of more than 40 service departments to provide citizens with more than 4,000 services. The Pangu e-Government Model connects a wide range of data sources, from charging piles, environment sensors, to gas meters, through a sensing network. Mass data from these sources, including texts, videos, and images, are collected and analyzed via multi-modal converged training combining natural language processing (NLP) and computer vision (CV). Pangu parses, understands, and dispatches issues to the responsible departments with handling suggestions provided for over 50% more efficient city governance.
In the automotive industry, one key challenge to autonomous driving is the seemingly endless possibility of corner cases. The Pangu Vehicle Model automatically generates samples for such complex driving scenarios using digital twins. This brings the over two weeks needed by an autonomous driving system to learn new corner cases down to under two days. To support autonomous driving for commercial vehicles used in coal mines, Pangu simulates various driving challenges, such as flying dust, long slopes, and steep turns, and automatically labels the generated samples. It takes only four person-months to adapt the model to new models of heavy trucks. Both Xinjiang Jiangna and Inner Mongolia Yimin's open-pit coal mines deploy Huawei's autonomous driving cloud service for commercial vehicles. Their error is now less than 0.2 m in lateral control and 0.1 m in parking for 60-ton trucks.
The Pangu Vehicle Model also supports automotive design, manufacturing, marketing, and R&D. Song Lei, Director of Digitalization Dept of FAW Jiefang, shared their experience with Huawei's Pangu Vehicle Model. This intelligent brain integrates deep into existing systems and serves as intelligent assistant for employees.
Song Lei, Director of Digitalization Dept of FAW Jiefang
In the meteorology domain, rainstorms are particularly difficult to predict. Worldwide, such storms and their resulting floods cause annual economic losses of hundreds of billions of dollars. Featuring a 3D earth-specific transformer (3DEST) architecture, Huawei's Pangu-Weather AI model is pre-trained on 40 years of global weather data and 10 years of precipitation satellite data. It generates 6-hour and 24-hour precipitation forecasts that are 20% more accurate than traditional model predictions.
On September 30, Huawei Cloud officially invites global users to test the Pangu-Weather model and its forecast capability. Huawei Cloud stands ready to work with customers and partners on localized precipitation forecast models for higher resolution. Rainstorm warnings are issued 24 hours in advance, instead of 3 hours.
Thailand has eight months of rain each year. To improve preparedness for extreme weather, the Thai Meteorological Department (TMD) plans to customize a Pangu weather model for the country. Dr. Chomparee Chompurat, Director General of TMD, and Zhang Ping'an, Executive Director of Huawei and CEO of Huawei Cloud, made a joint announcement at HUAWEI CONNECT 2023 of their innovation program for Thailand.
Dr. Chomparee Chompurat and Mr. Zhang Ping'an made a joint announcement of their innovation program.
In medicare scenarios, the Pangu Medical Model is trained on high-quality data from 16 million academic journals and equipped with over 1 million knowledge graphs. It works as an intelligent assistant for doctors and patients during medical examinations, diagnosis and treatment, and personal health management. In medical examinations, a missed or wrong diagnosis can be due to complex symptoms, too many test results, and insufficient doctor experience. Liangyi Xiaohui, a medical AI model Runda Medical developed on Pangu, is expected to be adopted by 4,000 hospitals for report interpretation and diagnosis, promoting inclusive healthcare.
Virtual human livestreaming is unleashing huge market potential in rural areas. Using the Pangu Virtual Human Model, farmers can quickly build their own virtual humans. With precise text and voice control, such virtual humans accurately and fluently introduce and sell products, interacting with viewers in real time in more than 20 languages, once trained in even one language. This model connects local products to a bigger market, boosting rural livelihoods.
For software R&D, the newly upgraded CodeArts Snap enhances code review and optimization. Connected to 23 development tools that are in the CodeArts pipeline, CodeArts Snap makes R&D 30% more efficient and helps developers write compliant and trustworthy code. In addition, the Pangu R&D Model connects to the Huawei Cloud Astro, our low-code platform, for a range of development scenarios, such as data dashboards, development workflows, and intelligent coding assistant. Developers automatically create applications by simply talking to AI.
According to Zhang Ping'an, Huawei Cloud has sent PhDs, experts, and scientists to the field over the past several years to help enterprises operationalize the latest technologies in construction sites, mines, and manufacturing workshops. So far, Huawei has helped more than 1,000 enterprises create real value from intelligent projects. This year, 200 PhDs are set to visit customer sites to bridge the gap between intelligent technology and industry needs, paving the way for customers to digital and intelligence.
In doing so, Huawei Cloud stays true to its mission statement of building AI for industries. Mr. Zhang said, adding that Pangu will reshape industries and reimagine work and life for every industry, every enterprise, and every individual.
Huawei Cloud Ascend AI Cloud Service Launched
In his speech, Mr. Zhang announced the launch of Huawei Cloud Ascend AI cloud service, available for purchase and use on the Huawei Cloud website from the day on. This cloud service offers a variety of ready-to-call AI models to boost wider AI adoption.
AI compute needs are skyrocketing in the full-intelligence era. Huawei estimates that by 2030, general-purpose computing power will increase by tenfold, and AI computing power will increase by 500-fold. This will be ever more demanding on data centers, especially self-built ones. Each server cabinet consumes six to eight times more energy to compute AI than other purposes, and requires a dedicated liquid cooling system. Foundation model training also can require hundreds of GPUs (sometimes even tens of thousands), a huge challenge for server connectivity, latency, reliability, and O&M.
Zhang Ping'an, Executive Director of Huawei and CEO of Huawei Cloud
To address these challenges, Huawei Cloud has built three AI compute centers in Gui'an, Ulanqab, and Wuhu. Enterprises have unlimited access to Ascend-powered AI compute without having to spend on GPUs or worry about supply continuity.
In addition, a dedicated zone for foundation models of Ascend AI cloud service is now officially available on the Huawei Cloud website. Enterprise users and developers can directly use mainstream open-source models such as LLaMA and GLM.
To ensure data security for foundation model training, Huawei Cloud adopts multiple E2E measures, including in-transit and at-rest encryption, data erasure, access control, and watermarking.
Huawei Cloud provides efficient, long-term stability and a complete tool chain for foundation model training. 1000-GPU training runs uninterrupted for 30 days, while tasks are restored in less than 30 minutes. With more than 30 visualization tools for data processing, model fine-tuning, and prompt engineering, foundation model development is now five times more efficient. It takes only one month, not five, to develop an industry-specific model with 100 billion parameters.
Building the Cloud Foundation for an Intelligent World with Systematic Innovation
Innovation is continuous. By investing in systematic futureproofing of core tech, Huawei Cloud aims to become the cloud of choice for the AI era.
In his speech, Bruno Zhang, Huawei Cloud CTO, predicts growth of 540 billion model parameters to number in the trillions with the advances of foundation models, multi-modalities, and MoE. This raises the bar for stable computing, large-scale parallel training, and overall architecture design. Challenges also exist in the rapidly growing dataset, storage, cleaning, and labeling, compounded by the massive token quantity. Given these challenges, foundation models and their applications require software/hardware system engineering of unprecedented complexity.
Bruno Zhang, CTO of Huawei Cloud
To address these challenges, Huawei Cloud embarks on systematic innovation of both software and hardware for AI.
At the system architecture layer, QingTian provides customers with hyperscale, high-performance AI compute. This peer-to-peer interconnected architecture builds on new high-speed network protocols to redefine compute, storage, and networking.
QingTian architecture supports Ascend AI cloud service (compute cluster), CANN (compute engine), MindSpore (AI development framework), and ModelArts (AI development platform) to optimize compute for developing and running foundation models and AI applications.
At the data processing layer, Huawei Cloud provides one-stop capabilities to locate, store, remember, and utilize enterprise foundation model data.
At the model training layer, Huawei Cloud Pangu models have a 5+N+X hierarchy that provides each industry and enterprise with custom models that best serve their own needs.
At the application development layer, Huawei Cloud integrates development pipelines, low-code/no-code platform, AI framework, and brand-new engineering suite to help customers quickly build AI native and foundation models into their applications.
Bruno Zhang, CTO of Huawei Cloud
These innovative services are deployed across clouds, networks, edges, and devices. Public cloud, edge cloud, and hybrid cloud are all available for deployment and implementation. If an enterprise needs a dedicated AI platform and versatile foundation models, Ascend AI cloud service and key foundation models work with existing data centers using Huawei Cloud Stack, a hybrid cloud platform. At the event, Bruno announced Huawei Cloud Stack 8.3 with mainstream availability by the end of September.
Bruno concluded his speech by reiterating Huawei Cloud's commitment to building the cloud foundation for AI with systematic innovation, and Huawei Cloud's position as a ubiquitous cloud foundation for an intelligent world.