At the ongoing Mobile World Congress (MWC2024), Huawei Cloud unveiled 10 AI technologies for integrating and enabling AI technologies. The objective is an AI-ready infrastructure tailored to each industry for a faster journey towards intelligence.
William Fang, Chief Product Officer at Huawei Cloud, emphasized the increasing demands on various aspects of cloud infrastructure due to the rapid advancement of AI and foundation models.
KooVerse: Huawei Cloud has 85 AZs in 30 Regions across over 170 countries and regions. This global cloud infrastructure covering compute, storage, networking, and security pushes latency down to 50 ms.
Distributed QingTian architecture: Foundation models require a 10-fold growth in demand for compute resources every 18 months, far surpassing Moore's Law. To address this challenge, this architecture evolved from the conventional primary/secondary one. Built on a high-speed interconnect bus (Unified Bus), QingTian surpasses the limitations in compute, storage, and networking for a top-class AI compute backbone with heterogeneous, peer-to-peer, full-mesh computing.
AI compute: Hyperscale and stable, AI Cloud Service supports trillion-parameter model training, and training jobs can run uninterrupted on a cluster over thousands of cards for 30 days, 90% of the time. Service downtime stays within 10 minutes. It provides over 100 Pangu model capability sets and 100 adapted open source large models out of the box.
AI-Native storage: Training models needs mountains of data, and Huawei Cloud handles this demand with a three-pronged approach: EMS memory service stores petabytes of parameters with 220 TB ultra-large bandwidth and ultra-low latency down to the microsecond; SFS Turbo cache service for high throughput and concurrency of tens of millions IOPS enables warm-up of 1 billion data records in just 5 hours, not 100; Object Storage Service (OBS) knowledge lake reduces 30% costs in storing training and inference data.
E2E security: The full lifecycle covers model runtime environments, training data, the models themselves, generated content, and applications. This ensures robust, secure, and compliant models and applications.
GaussDB: This next-generation database features high availability, security, performance, flexibility, and intelligence, as well as simple and smart deployment and migration. Specifically, its enterprise-class distributed architecture ensures high availability thanks to zero intra-city dual-cluster RPO, complete isolation of software and hardware faults, and zero service downtime. For security, it is certified CC EAL4+, the highest level in the industry. For automation, GaussDB enhances database migration, deployment, and migration as the world's first AI-native database.
Data-AI convergence: The explosion of foundation models means "Data+AI" is now "Data4AI and AI4Data". Huawei Cloud LakeFormation unifies data lake from multiple lakes or warehouses so one copy of data is shared among multiple data analytics engines and AI engines without data migration. Three collaborative pipelines — DataArts, ModelArts, and CodeArts — then orchestrate and schedule data and AI workflows. They drive online model training and inference with real-time data. The AI4Data engine makes data governance more intelligent, from data integration, development, to quality and asset management.
Media infrastructure: In this AIGC and 3D Internet era, Huawei Cloud has built a media infrastructure of efficiency, experience, and evolution. Jamy Lyu, President of Huawei Cloud Media Services, shared how Huawei Cloud has innovated and integrated media services into a wide range of industry-tailored solutions. For efficiency, Huawei Cloud MetaStudio, the content production pipeline that include Workspace and AIGC-based virtual humans, generates content more quickly and better. For experience, Huawei Cloud Live, Low Latency Live, and SparkRTC empower more seamless live experiences. For evolution, Huawei Cloud provides AIGC and 3D space services with real-time user interaction. All these combine to boost the business and user experience to the next level.
Landing Zone: Enterprises use and manage resources better on Huawei Cloud thanks to unified account, identity, permissions, network, compliance, and cost management. Now multi-tenancy and collaboration are seamless among personnel, finance, resources, permissions, and security compliance.
Flexible deployment: All mentioned Pangu model capabilities and services can work in public cloud, dedicated cloud, or hybrid cloud. For example, customers can build and run dedicated AI platform and foundation models in their existing data centers using Huawei Cloud Stack, a hybrid cloud solution.