RoCE's penetration will effectively stimulate the demand growth and product technology iteration of the AI computing interconnect industry chain. Headquartered manufacturers are expected to maintain shipment growth with more solid market share, while domestic chip manufacturers are expected to achieve technological and customer breakthroughs.
According to Zhongtong Finance, Soochow Securities released a research report saying that IB network was dominant in the early stage of AI computing construction, but the industry has been exploring the possibility of using Ethernet to adapt to AI computing. The Ultra Ethernet Consortium (UEC) emerged, and network, system, and cloud computing manufacturers in various links of the industry chain such as Broadcom, Nvidia, and Meta have continued to invest and make progress. RoCE is expected to gradually replace IB as the mainstream. RoCE's penetration will effectively stimulate the demand growth and product technology iteration of the AI computing interconnect industry chain. Headquartered manufacturers are expected to maintain shipment growth with more solid market share, while domestic chip manufacturers are expected to achieve technological and customer breakthroughs.
Why use Ethernet & RoCE? Compared with traditional TCP / IP technology, RDMA is more in line with the requirements of high concurrency and low latency of AI computing, so it is a better choice for network technology. IB and Ethernet both support RDMA, but IB naturally supports RDMA and is the local optimal solution for short-term rapid, high-quality, and high-volume implementation of computing power on AI computing construction in the early stage. Ethernet has a deep application foundation and low cost, and is expected to be the best solution in the future.
How does Ethernet achieve high requirements for AI interconnection? AI interconnection mainly faces two major problems: 1) the long-tail effect brought about by the significant increase in "elephant flows" can be solved through RoCE's adaptive routing function; 2) congestion caused by multiple data receivers sharing the same receiving end in different computing processes can be solved through RoCE's switch congestion control algorithm + buffer pool.
What industrial changes does Ethernet bring? 1) Switch capacity is increased, and RoCE supporting functions such as adaptive routing and congestion control are added. At the same time, more abundant software and hardware also provide greater space for white-box switches, which is conducive to their further penetration; 2) The growth in inferring demand has opened up the implementation of RoCE. Cloud factories have accelerated self-built inference computing power to bring about 800G optical module additions, while Nvidia's customer structure continues to optimize, based on the advantages of training and inference on one platform leading the iteration and application of cutting-edge products such as 1.6T; 3) Silicon photonics has a three-layer industrial logic of guaranteeing optical module supply, undertaking the trend of silicon-based co-packaging, and reducing costs. It is expected to accelerate penetration.
Related symbols: Recommended Zhongji Innolight (300308.SZ) and Suzhou TFC Optical Communication (300394.SZ). Suggest paying attention to Eoptolink Technology Inc. (300502.SZ), Source Photonics (688498.SH), and Eoptolink Communication (688702.SH).
Risk warning: downstream demand is lower than expected; customer development and share are lower than expected; product research and development, production and quantity are lower than expected; industry competition intensifies.