【英文】摩根士丹利报告：全球半导体竞赛：中国将如何布局（35页）

In the past, CPUs have been the biggest factor affecting the performance and speed of datacenters. However, if there are bandwidth  limitations, even the most powerful CPU cannot deal with increasing  M data traffic. To achieve faster networking speeds, then, a more  diverse range of companies needs to work together, including semiconductor manufacturers, hardware players, and telecom operators.  Semiconductor companies, for example, need to come out with more  switch ICs to deal with internet traffic. However, CPUs are approaching certain limits as their computing  power cannot fully match the requirements of big data analysis. As  a result, GPUs have emerged as an alternative to be used in parallel  computing for AI training. Networking speeds continue to rise but  data transmission within the system has reached a limit. This implies  that interfaces in the system also need to be upgraded to speed up  the whole datacenter. Both I/O and communications interfaces need  to be upgraded.

Localizing datacenter semiconductors is a tough task because of the  high technology entry barriers. For example, Hisilicon has introduced  its own CPU for Huawei servers, and government also wants to fully  support it and increase the penetration. But given the performance  gap with Intel, the pace of increasing penetration will be gradual. The  power management IC sector is also dominated by foreign companies like Infineon and TI. However, high-speed interfaces are now presenting great opportunities for Greater China semi companies. Right now most high-speed  interface chips are still dominated by US companies. However, we  believe the performance gap between Greater China semis and their  US peers is small. Some companies in Asia have even better design  capabilities than US companies. Huawei, Inspur, and Dawning  account for 20-30% of global server demand and offer significant  opportunities for Greater China semi companies.