Investment Rating - The report provides a positive investment rating for the electronic industry, indicating a strong growth outlook driven by advancements in AI and GPU cluster technologies [13]. Core Insights - The growth trend of GPU clusters is significant, with expectations for exponential growth in cluster size due to the need for efficient training of large models. NVIDIA plans to interconnect over 10,000 GPUs in 2024, over 100,000 in 2025, and over 1 million in 2026 [2]. - The expansion of GPU clusters will increase the number of network layers, enhancing the ratio of switches to GPUs. The report highlights the use of Fat-Tree architecture in data center networks to meet high bandwidth and low latency requirements [2]. - The report anticipates a shift towards RoCE protocol in large clusters as InfiniBand's interconnect limit is reached, with RoCE offering better compatibility and distributed networking [2]. - The Ethernet switch market is expected to see significant growth, with projections indicating an increase from over $20 billion in 2023 to nearly $30 billion by 2026, driven by rising interconnect speeds and demand for high-performance computing [2]. Summary by Sections GPU Cluster Growth - GPU cluster sizes are expected to grow exponentially to meet the demands of large model training, with NVIDIA's roadmap indicating substantial increases in interconnected GPUs over the next few years [2]. - The report details the maximum number of GPUs that can be interconnected based on different network architectures, highlighting the scalability of GPU clusters [6]. Network Architecture - The report discusses the advantages of using CLOS network architecture for GPU interconnects, emphasizing the need for high bandwidth and low latency in data centers [2][7]. - It compares InfiniBand and RoCE protocols, noting the lower latency of InfiniBand but the potential for RoCE to become more prevalent as cluster sizes increase [7]. Market Opportunities - The report identifies key players in the Ethernet switch market, such as Huadian Technology and Marvell, which are positioned to benefit from the anticipated market growth [2]. - It highlights the expected increase in demand for high-capacity Ethernet switches, driven by the need for efficient data transfer in AI applications [2].
“AI的裂变时刻”系列13:计算集群规模提升趋势明确,以太网交换机市场空间广阔
GF SECURITIES·2024-07-04 02:31