从 Chiplet 到超节点：奇异摩尔正在塑造中国 AI 算力的“互联底座”

Core Viewpoint - The competition in computing power is shifting from single GPU performance to cluster-based systems, emphasizing the importance of chiplet, packaging, and interconnect technologies as the foundation of the new computing ecosystem [1][3]. Group 1: Interconnect as a Key Component - Interconnect and scheduling are essential for realizing the full potential of cluster computing, as AI models demand higher computational scales [3][4]. - The current state of domestic AI network cards lags behind international standards, with the most advanced being 200G compared to 800G or even 1.6T abroad, highlighting the need for significant improvements in interconnect architecture [3][4]. Group 2: Challenges of Supernode Architecture - The emergence of supernode architecture is reshaping system models and presents significant challenges, particularly in achieving high bandwidth and low latency interconnects [4][5]. - The lack of established standards for supernodes has led to a proliferation of protocols, indicating a diverse and evolving landscape in this area [5][6]. Group 3: Chiplet Implementation and Standardization - The supernode chiplet concept is seen as a solution to the challenges of chiplet implementation, allowing for shared R&D costs and faster product development [6][8]. - The HPDE architecture enables programmability and compatibility with multiple standards, facilitating the transition to unified supernode standards [6][8]. Group 4: Importance of Heterogeneous Computing - The future of computing may involve heterogeneous integration at the chip level, necessitating high interconnect bandwidth and effective task scheduling across different architectures [7][8]. - Both interconnect and scheduling are critical for optimizing performance, as they are interdependent in the context of heterogeneous computing [7][8]. Group 5: Industry Trends and Product Development - The AI sector is witnessing increased participation from various companies, indicating a collective shift towards AI-focused solutions across the semiconductor industry [12][14]. - The company has developed a comprehensive product lineup, including high-performance interconnect chiplets and AI-native network products, aimed at establishing a robust interconnect infrastructure for the AI computing industry [12][14]. Group 6: Future Goals and Market Positioning - The company aims to collaborate with clients to integrate new technologies into their products and expand into larger markets over the next one to two years [14][15]. - The focus on interconnect solutions positions the company as a key player in the evolving landscape of AI networking, striving for performance breakthroughs and industry standardization [15].