CICC: Google leads ASICS self-development to accelerate the redefinition of hardware value different from GPGPU architectures

The Zhitong Finance App learned that CICC released a research report saying that the launch of Google (GOOGL.US) Google TPUv7 marks the ASICS cluster's acceleration of self-research on an architecture different from traditional GPGPU, bringing heterogeneity and reshaping to hardware value. At the same time, it is expected to accelerate the rapid rise in the volume and price of AI computing power hardware such as PCBs, liquid cooling, and power supplies. Looking ahead to 2027, the AI PCB/liquid cooling/power chip market size is expected to reach US$216.5/201.8/18.39 billion, respectively. It is recommended to focus on Shennan Circuit (002916.SZ), Shengyi Technology (600183.SH), Dongshan Precision (002384.SZ), Pengding Holdings (002938.SZ), Industrial Fulian (601138.SH), Hongteng Precision (06088), and Azure Lithium Cell (002245.SZ).

CICC's main views are as follows:

The 10-year evolution of Google's TPU architecture

Since Google officially disclosed the TPU v1 in 2016, it has gone through ten years of architectural evolution. The TPU has evolved from a pulsating array dedicated to inference to a training chip for nearly 10,000 card clusters, which introduced an OCS optical switching architecture and HBM high-bandwidth storage. At present, Google has released the TPUv7 chip, which has further broken through the dual-core packaging architecture, and the linear acceleration ratio in hyperscale clusters has improved significantly.

Google's next-generation TPUv7 hardware side brings big changes

The tray architecture includes 16 standardized computing trays, each carrying 4 TPU chips; the power supply architecture uses a +/- 400V high voltage DC scheme (HVDC); in terms of server cooling, a 100% liquid cooling architecture is used, using a large cold plate design, covering 4 TPUs and VRMs; the maximum cluster size supports 144 rack interconnections, or 9216 TPU chip clusters.

The bank dismantled the value of Google's TPUv7 cabinet solution. Among them, the value of TPU, PCB, liquid cooling, power supply, cable, etc. was 54.4/4/7/7.1/0.4 million US dollars, totaling about 730,000 US dollars. The bank believes that with the gradual increase in TPU shipments and product structure iterations, demand for the hardware side is expected to increase rapidly, which is expected to drive the PCB/liquid cooling/power chip market volume and price increase rapidly. According to Google's procurement caliber, the AI PCB/liquid cooling/power chip market size is expected to reach 36.9/6.06/3.1 billion US dollars in 2027; judging from the overall GPU+ASICS procurement demand, the AI PCB/liquid cooling/power chip market size is expected to reach 216.5/201.8/18.39 billion US dollars respectively in 2027.

risk

The demand for AI and computing power infrastructure falls short of expectations, the risk of increased competition in the industry, and the risk of a shortage of key devices.