浪潮信息：AIGC大模型算力平台参考设计_20页_5mb

Open Accelerator AI Server Design Guide Summary

Background

The Open Accelerator (OAI) initiative was developed by the Open Compute Project to address challenges in AI acceleration, specifically for generative AI. With the rise of large models like GPT and others requiring massive compute resources, OAI promotes standardized open specifications for AI servers. It aims to enhance compatibility, efficiency, and scalability in AI infrastructure.

Design Principles

• 4.1.1 Application-Oriented: Focuses on real-world AI application scenarios to ensure designs meet practical needs.
• 4.1.2 Diverse and Open: Supports adaptation to various AI acceleration technologies through open standards.
• 4.1.3 Green and Efficient: Prioritizes power efficiency and advanced cooling solutions to handle high energy consumption from AI chips.
• 4.1.4 Comprehensive: Emphasizes scalable planning and efficient resource management for large AI models.

Design Guidelines

• 4.2.1 Multi-Dimensional Collaborative Design: Involves integrated design from node to cluster levels, including system architecture, OAM modules, UBB baseboards, hardware, thermal, and management aspects. For example, UBB supports 8 OAM modules with high-speed interconnects and Power over Ethernet.
• 4.2.2 Comprehensive System Testing: Covers structural, thermal, stability (e.g., stress and endurance tests), and compatibility (e.g., OS, frameworks, models) to ensure reliability and functionality.
• 4.2.3 Performance Measurement and Optimization: Tests compute performance, interconnection bandwidth, and uses tools like RDMA for optimization in AI frameworks and models.

Key Components

• OAM Modules: Define standard AI accelerator interfaces, enabling Python code or OAM DPDK for low-latency communication.
• UBB Baseboards: Support 8 OAM modules, facilitating high-density configurations with topologies like FC and HCM for efficient compute scaling.

Performance Considerations

Performance assessment includes basic and interconnection tests (e.g., PCIe and P2P bandwidth), with additional focus on model-specific performance tuning (e.g., TensorFlow, PyTorch). Strategies for optimization involve handle-level communication, efficient management, and resource allocation.