CICC released a research report saying that liquid cooling, as a method with higher heat dissipation efficiency, is expected to replace traditional air cooling with the advantages of increasing computing power deployment density and reducing system power consumption, and achieve rapid growth along with the increase in the scale of AI computing power.
The Zhitong Finance App learned that on March 19, CICC released a research report saying that on March 19, Nvidia held a GTC conference and unveiled the first Blackwell series GPU, GB200 superchip, and the GB200 NLV72, a rack-level system based on this. The cooling method changed from air cooling to liquid cooling. As a method with higher heat dissipation efficiency, liquid cooling is expected to replace traditional air cooling with the advantages of increasing computing power deployment density and reducing system power consumption, and achieve rapid growth along with the increase in the scale of AI computing power. It is recommended to focus on liquid cooling-related targets: Shuguang Digital Innovation (872808.BJ), Gaolan (300499.SZ), Ziguang (000938.SZ), Lenovo Group (00992), Industrial-Fulian (601138.SH), and Nvidia (NVDA.US).
The main views of CICC are as follows:
Liquid-cooled heat dissipation enables high computing density, increased penetration rate is the general trend, and the industry is being implemented at an accelerated pace
CICC believes that Nvidia introduced liquid cooling and heat dissipation into the computing node to enable high computational density. The GB200 NVL72 released at the GTC conference can achieve 720 PetaFlops AI training performance and 1440 petaflops inference performance. It consists of 18 trays, each containing 2 GraceCPUs and 4 Blackwell GPUs, and is equipped with liquid cooling; in addition, manufacturers such as Lenovo Group, Ziguang, and Shuguang Digital Innovation are also continuing to enrich their product matrices and accelerate the implementation of liquid cooling solutions. The introduction of Nvidia's liquid cooling is expected to be a milestone, confirming the superiority of liquid cooling technology while also raising the market's expectations for the pace of liquid cooling implementation and potential market space.
Liquid cooling has the advantage of high heat dissipation efficiency, which helps to increase the power density of a single cabinet
Compared with traditional air cooling, the advantages of liquid cooling are: 1) In terms of heat dissipation efficiency, according to CDI consultant data, cooling equipment accounts for 33% of the total energy consumption of data centers. Liquid cooling directly absorbs and takes away the heat generated by computing power hardware such as GPUs through liquid, which can improve cooling efficiency and reduce total energy consumption; 2) In terms of computing density, since liquid cooling does not require reserving a large amount of space for fans, etc., it is possible to deploy high-density computing units to reduce floor space. In addition, a more compact hardware layout is also conducive to shortening the physical distance between computing units and reducing signal transmission delays .
According to Nvidia GTC, taking the GPT-MoE model with a 90-day training parameter of 1.8 tons as an example, if the H100 is used, 8,000 cards are required, and the power consumption is 15MW; if the GB200 NVL72 is used, only 2,000 cards are required, and the power consumption is 4MW; based on this, CCC estimates that the power consumption corresponding to the GB200 NVL72 single cabinet exceeds 100KW. CICC believes that with the advantages of high heat dissipation efficiency and increased computing power deployment density, liquid cooling is expected to achieve rapid growth driven by the increase in AI computing power scale and liquid cooling penetration rate.
risks
The iteration and cost reduction of computing power chip products fell short of expectations; the penetration rate of liquid cooling fell short of expectations.