Cloud Computing in the AI-Native Era: Baidu Cloud's AI-Native Architecture and Latest Product Innovations

This article is a keynote speech by Baidu Vice President Xie Guangjun at the 2023 Baidu Cloud Intelligence Conference, titled "Cloud Computing in the AI-Native Era." The presentation outlines Baidu's comprehensive AI-native cloud architecture and its latest product advancements.

Computing Infrastructure: Baidu launched the 7th generation cloud server instance G7 featuring the latest Intel EMR processors, delivering 10% improved performance. Two domestic AI computing instances were introduced: the new Kunlun R300 bare metal server, offering 50% performance improvement in large model inference scenarios, and the Ascend 910B-based elastic high-performance computing instance, providing 40% improvement in large model training. All instances support the second-generation Baidu DPU network cards.

Network Innovation: Baidu developed a new self-developed gateway platform integrating X86 CPU, programmable switching chips, and FPGA acceleration cards, achieving T-level traffic forwarding capability with average latency reduced by over 20 times.

Storage Upgrades: Baidu's unified storage technology foundation includes metadata storage unified to TafDB and a unified data foundation called Aries supporting multiple data models. The storage products include object storage BOS with flat/hierarchical namespace migration, block storage CDS with microsecond-level latency, and parallel file storage PFS with extreme performance variants.

AI Infrastructure - Baidu Baige 3.0: Specifically optimized for large models, Baige 3.0 offers training and inference acceleration tools, high-performance communication libraries, and large image distribution acceleration. RDMA bandwidth effectiveness reaches 95%, with throughput improved by 30-60%. The platform provides cluster fault detection, automatic fault tolerance, and Flash Checkpoint functionality, achieving over 98% effective training time for 10,000-card level tasks.

Intelligent Computing Network Platform: This platform connects distributed heterogeneous computing resources including intelligent computing centers, supercomputing centers, and edge nodes, forming a unified computing network resource pool with intelligent scheduling capabilities.

Database and Big Data: GaiaDB 4.0 features parallel queries achieving 10x performance improvement, columnar storage indexes, and 60% overall performance improvement. The Database Intelligent Cockpit uses large model capabilities for automated database optimization. The BMR Spark 3.2 engine delivers 2x performance improvement over community versions. Sugar BI's intelligent data query (SugarBot) enables natural language-based data analysis, reducing steps from 6 to 3 and improving efficiency by 5 times.

Distributed Cloud: Includes edge computing nodes (BEC), local computing clusters (LCC), and private cloud ABC Stack with integrated Qianfan large model platform.

Application Platforms: Intelligent Video Cloud Platform 4.0 provides one-stop AI-powered audio/video solutions including AIGC intelligent highlights, intelligent video enhancement, and digital watermarking. The low-code platform Aisuda integrates large model capabilities for rapid application development.