AI server: amid the surge of liquid cooling don't count out air cooling fans yet

Aaron Lee, Taipei; Jerry Chen, DIGITIMES Asia 0


Liquid cooling technology has been gaining attention due to the substantial heat dissipation requirements of AI servers. But does this mean that air-cooling fans are becoming obsolete? Fan manufacturer Sunon rebuts this presumption, asserting that it is, in fact, a misconception.

Sunon stresses that air cooling technology is also advancing, with fans capable of directly dissipating heat expected to increase from 500 to 800 watts. Regardless of whether it's water-cooling or liquid-cooling, the core objective of heat dissipation remains constant. Fans merely shift from the front end to the back end, with a subsequent increase in their unit price as they transition.

Air cooling fans: far from obsolete

AI servers, which require GPUs or ASIC chips, have significantly higher heat dissipation coefficients, leading many to question whether the existing air-cooling systems used for servers have reached their heat dissipation limits. As a result, speculation looms about the potential replacement of the air-cooling solutions with water or liquid-cooling solutions. Nonetheless, this perspective is broadly rejected by server manufacturers, cooling module producers, and fan manufacturers as an oversimplification.

Firstly, servers encompass not just AI servers but also general-purpose ones. Although the demand for AI servers is on the rise, they represent a smaller share. Most servers are general-purpose servers, and they overwhelmingly rely on air cooling due to cost-effectiveness, reliability, the maturity of this technology, and the robustness of the supply chain.

Secondly, aside from newly built data centers, existing data centers primarily employ air-cooling designs and have not considered the pipeline and spatial configurations necessary for water cooling or liquid cooling. It remains unclear how many newly built data centers will adopt liquid-cooling designs. Charles Liang, CEO of Supermicro, has predicted that the penetration rate of liquid cooling in data centers will reach 20% in "the coming years."

However, the exact duration constituting "the coming years" Liang referred to remains unpredictable. Even within this timeframe, only 20% will potentially adopt liquid cooling, signifying that the majority, 80%, will continue to use air cooling. Shen Ching-Hang, Chairman of Asia Vital Components (AVC), a fan and cooling module manufacturer, mentioned that it will likely take years until after 2025 for liquid cooling to gradually gain more prominence.

The slow rise of liquid cooling

Nonetheless, both server manufacturers and cooling companies are not ignoring the trend of water cooling and liquid cooling. Supermicro states that it's actively developing data center solutions for liquid cooling. Companies like Quanta, Wiwynn, and Inventec are also introducing liquid-cooling solutions. Meanwhile, companies like AVC, Sunon, Auras, and Taisol are exploring various options, including air cooling and liquid cooling.

Sunon notes that while customers are considering the adoption of solutions like water-cooled boards, this doesn't mean fans will lose their role. Instead, fans shift from the heat source to the rear, as energy isn't lost; heat needs to be efficiently conveyed to the rear. This applies whether it's a water-to-air or water-to-water configuration, as fans continue to play a pivotal role in effectively expelling heat.

Even in immersion cooling, pump motors are necessary where liquid flows within the cabinet. Fan manufacturers already possess the technology and products for motors. Hence, irrespective of the solution chosen, it falls within the purview of fan manufacturers. Sunon states that whether it's water-cooled boards, CDUs, and more, all of them are in the process of development and testing.

The challenge of larger fans

Moreover, Sunon further points out that if customers opt for water-cooling or liquid-cooling solutions, the fan size will need to be larger than those used in air cooling. This not only implies different electronic components but also different communication protocols. For instance, the voltage may increase from the original 12 or 48 volts to 380 volts, resulting in a three to five-fold increase in technical difficulty. In terms of space, a 4U AI server chassis may utilize 8cm fans, while water-to-air fans could be 20cm, tripling in size.

Sunon argues that larger fans will naturally have higher average unit prices even though the increase is not directly proportional. Factors such as power consumption need to be considered. However, larger fans encounter less competition due to fewer players in this field. Additionally, patents also raise the barriers for market entry, even if invisibly.