What is the difference between liquid-cooled data centers and air-cooled cooling technology?

The choice of cooling method for data centers in the next stage is not decided by the data center temperature control manufacturers, nor by the builders and operators of data centers. The core driving factors are jointly decided by the downstream internet/cloud computing/AIGC industries (such as OpenAI, Meta, and Baidu) and the manufacturers of computing power/chips/servers (such as Nvidia and Huawei).

What is clear at present is that the market growth rate of the internet/cloud computing industry in China and even globally has slowed down (there are signs of reaching a peak in terms of user scale and commercialization), while the AIGC industry is experiencing a strong breakthrough. The large-scale training and inference of AIGC requires the adoption of new computing power chip servers and a new IT architecture, and also brings higher heat dissipation requirements.

From the perspective of industrial demand, the overseas market has benefited from the development of AIGC represented by ChatGPT, driving the rapid construction of overseas computing power centers; in China, due to the limitations of high-performance chip imports and the maturity and application scenarios of domestic large models, the construction of computing power centers in China has begun to be implemented in stages, but the large-scale construction and implementation of leading manufacturers has not yet been fully rolled out. Overall, data centers serving AIGC, particularly large ones for training large models, will be the main driving force for the scaled construction of large-scale computing centers in the initial stage. And the mainstream solutions for these large computing centers are mostly liquid cooling with cold plates. Therefore, the next stage's incremental market for domestic data centers will indeed be represented by the mainstream demand for large-scale computing center construction and liquid cooling solutions and products.

From a technical perspective, due to the limitations of air cooling in terms of heat dissipation density and efficiency, the latest mainstream high-performance computing servers in the industry have already fully switched to liquid cooling solutions. In the foreseeable future, with the rapid growth of the AIGC industry, the demand for chip computing power will also grow exponentially. During this process, air cooling will no longer be the choice for high-performance computing servers, and from a technical perspective, unless the heat effect of computing chips is reversed (such as whether quantum computing can achieve lower energy consumption and heat dissipation per unit computing power), there is no doubt that the cooling method for the main equipment (mainly IT servers) in the incremental market for data centers will shift from air cooling to liquid cooling. The current mainstream liquid cooling solution for data centers is the cold plate liquid cooling, which is determined by the server manufacturers (such as Nvidia, Supermicro, and Inspur) for AI computing servers (which are designed to match the cold plate liquid cooling form). However, the technical solution of cold plate liquid cooling naturally brings about the demand for "wind-liquid fusion": The liquid cooling backplane can only remove heat from the CPU and GPU, while other devices in the server, including hard drives, memory, power supplies, and IC devices, also need to be cooled. These devices cannot be individually equipped with liquid cooling backplanes, and heat must be dissipated through the air-side cooling form. This also gives rise to the inherent architecture of the cold plate liquid cooling solution - "wind-liquid fusion". This means that in the cooling architecture of a new AIGC computing center, both liquid cooling cycles and supporting equipment systems for chip (CPU and GPU) cooling are needed, as well as traditional air-side cooling temperature control equipment. The current industry's generally accepted ratio of "wind-to-liquid" (air-side cooling demand to liquid-side cooling demand) is around 3:7 to 4:6. Therefore, in an AIGC-driven computing center, the cold plate liquid cooling solution actually brings the industry a combination of liquid cooling and air cooling (i.e., "wind-liquid fusion") demand. It is worth noting that the demand for AIGC-related AI supercomputing centers includes the training, inference, and some computing power scenarios of small models, and the single rack load density and air-liquid ratio of these scenarios are different. There are also some statements in the industry that the current cloud computing data center architecture (the main cooling solution is air cooling) can also adapt to the computing power demand of large-scale model training, and even in the era of AIGC fully arriving, we still have the demand for large-scale general computing power center construction. Therefore, when Chinese data center construction parties plan for new computing power centers, in the uncertain prospects and pace of AIGC development, they should consider both general computing power demand and large model inference demand, as well as large model training scenarios. Under this background, adopting large-scale wind-liquid fusion architecture and reserving flexible air-liquid ratios to meet future business extension and expansion needs has become the inevitable choice of data center construction parties. It can be said that the coming era is the era of "wind-liquid fusion".

Perhaps in the near future, with the further evolution of the AIGC industry and its extreme squeezing of computing power and performance, the current industry's main cold plate liquid cooling will also be unable to bear its heat dissipation demand. At this point, the market will force server manufacturers to choose higher-efficiency cooling solutions such as "immersion liquid cooling." Once this era fully arrives, the "wind-liquid fusion" problem of existing cold plate liquid cooling will be resolved. Because the entire server is fully immersed in the cooling liquid, there is no need to consider which parts need to be cooled by cold plates and which parts need to be cooled by air side. At this point, the data center industry can be said to have truly bid farewell to the "pure air cooling" and "wind-liquid fusion" era and ushered in the "pure liquid cooling" era.

The last sentence summarizes: For data center cooling solutions, the "pure air cooling" era is coming to an end, and the "wind-liquid fusion" era is racing towards us; from a distance, we can already catch a glimpse of the masts of the "pure liquid cooling" era on the horizon.