液冷成为2023年DCW展会上热门话题
原创 江秋健 DeepKnowledge 2023-09-22 17:30 发表于河北
收录于合集 #数据中心  #DCW 展会  #液冷 #直接到芯片  #浸没式

液冷成为2023年DCW展会上热门话题
Liquid cooling a hot topic at Datacenter World 2023

Analysts - Perkins Liu Tuesday, May 30 2023

译 者 说
2023年奥斯汀的DataCenter World大会上，液冷成为人们谈论最热的话题之一，液冷已经逐步走向部署和应用。

Introduction
介绍

液冷是最近在德克萨斯州奥斯汀举行的Data Center World 2023活动上讨论的热门话题之一。来自数据中心生态系统各个领域的从 业人员分享了关于动态数据中心市场液冷的驱动因素、挑战、成功、想法、实践和愿景。在展览现场，液冷技术和解决方案公司展示了各种产品和方法，应用于不同的场景以应对不断增长的功率密度和可持续性挑战，其中包括直接芯片冷却技术(DtC)和单相、两相浸没式冷却技术。

Liquid cooling was one of the hotter topics of discussion at the recent Data Center World 2023 event in Austin, Texas. Practitioners from all fields in the datacenter ecosystem shared drivers, challenges, successes, ideas, practices and visions about liquid cooling in the dynamic datacenter market. On the exhibition floor, liquid cooling technology and solution companies demonstrated a variety of products and approaches, including direct-to-chip (DtC) and immersion, single phase and two phase, in different applications to address ever-increasing power density and sustainability challenges.

几十年来，数据中心一直使用空气冷却，使用不同的气流管理方法，如列间和通道密封(热和冷)。行业已经看到风冷系统从房间到行间到机架水平，不断提高精度和效率，以解决由芯片设计和高性能计算(HPC)需求驱动的日益增长的机架功率密度。热设计功率(TDP)是芯片功率密度的一个关键指标，目前正在缓慢地向200-300瓦的方向发展，而大多数数据中心仍然采用空气冷却。液体冷却一直局限于超级计算、HPC和加密采矿等应用，这些应用可能需要更高的密度。

Datacenters have been cooled by air for decades, using different airflow management approaches such as in-row and aisle containment (hot and cold). The industry has seen air cooling systems move from room to row to rack level, for continuously improved precision and efficiency to address increasing rack power density driven by chip design and high-performance computing (HPC) requirements. Thermal design power (TDP), a key indicator of chip power density, has been slowly marching toward 200-300 watts, with most datacenters still comfortably cooled by air. Liquid cooling has been limited to applications such as supercomputing, HPC and crypto mining, which can require a higher density.

然而，近年来，随着人工智能(AI)、机器学习(ML)、数据分析(data analytics)、高性能计算(HPC)和物联网(IoT)等领域的进步，芯片技术的快速发展加速了部分芯片(如CPU/GPU)的密度，在两年内进入700瓦甚至1000瓦的范围。这可能会给目前使用空气冷却的数据中心带来压力，并导致更多地使用液体冷却。除了功率密度方面的挑战，由于液体冷却的一些好处，围绕可持续性的日益增长的担忧和举措可能会进一步加快在数据中心采用液体冷却;例如，减少或取消风扇，提高电力使用效率(PUE)，减少水的使用，节省空间和增加热量再利用的机会。

Recently, however, progress in artificial intelligence (AI), machine learning (ML), data analytics, HPC and internet of things (IoT) has been accompanied by the rapid evolution of chip technology, which has accelerated the density of some chips (e.g., CPU/GPU), entering the range of 700 watts and potentially 1,000 watts in two years. This may put pressure on datacenters currently using air cooling and lead to more use of liquid cooling. In addition to power density challenges, growing concerns and initiatives around sustainability could further accelerate the adoption of liquid cooling in datacenters, thanks to some of liquid cooling's benefits; e.g., reduced or eliminated fans, improved power usage effectiveness (PUE), reduced water use, saved space and enhanced opportunities for heat reuse.

从风冷到液冷的转变不仅是技术上的考虑，也是法规、供应链和经济上的考虑，甚至是文化上的转变。这可能是一个漫长的过程，并不是所有的风冷数据中心都将或应该被液冷取代。在小组讨论中，IT服务提供商、芯片、服务器和数据中心基础设施制造商、冷却技术初创公司以及数据中心设计和建造公司都很乐意深入探讨液冷这个话题，并提供了深入而全面的观点。人们普遍认为，在许多情况下，液体冷却比空气冷却更好地应对芯片密度上升带来的物理挑战。

The transition from air cooling to liquid cooling is not only a technological consideration, but also a regulation, supply chain and economic consideration, even a culture shift. It will likely be a long process and not all air-cooled datacenters will or should be replaced by liquid cooling. In a panel discussion, IT service providers, manufacturers of chip, server and datacenter infrastructure, cooling technology startups and datacenter design and build firms were happy to dive into the topic of liquid cooling, offering an in-depth and holistic view. It was generally agreed that liquid cooling is, in many cases, better than air cooling at dealing with the physical challenges created by rising chip densities.

这种不断上升的密度通常源于AI、ML、HPC和其他计算创新所需的处理器性能增强。当芯片TDP进入200到300Wh的范围时，需要特殊的气流管理，以及增加风扇的体积和速度，足以快速带走热量。这也是液体冷却开始变得更有效的地方。当TDP达到350-400Wh时，使用空气冷却服务器变得更加困难和昂贵。换算成机架密度，每机架20-50kW大致是风冷和液冷的重叠范围。TDP的持续上升可以很容易地推动机架密度超过50kW甚至进入三位数范围，而液体冷却是最有效的方法(在某些情况下，唯一的方法)来对这种密度的机架进行冷却。

This rising density often results from the enhanced processor performance required for AI, ML, HPC and other computing innovations. When chip TDP enters the 200- to 300-watt range, special airflow management, together with increased fan volume and velocity, are required to take the heat away quickly enough. This is also where liquid cooling starts to become more effective. When the TDP reaches 350-400 watts, it becomes much more difficult and costly to cool the server using air. Translated into rack density, 20-50 kW per rack is roughly the overlap range of air cooling and liquid cooling. The continued rise in TDP could easily push the rack density over 50 kW or even into the three-digit range, and liquid cooling is the most efficient way (in some cases, the only way) to cool such densities.

图片
数据中心国际专家培训

图片 ATD设计课程
没有做过运维的设计师就无法设计出好用的数据中心吗？ATD课程将彻底解决这个问题

图片 AOS运维课程
运维管理的直接目标是优秀的运维人员而非设备本身；全球权威的AOS运维管理专家课程将带您深刻理解运维管理的本质

图片 ATS管理课程
正确和系统地了解Tier分级体系会提升数据中心的项目投资回报、减少业务中断风险，ATS课程将全面带您学习Tier知识，帮助您有效提升企业的运营指标和内外部客户的满意度。

点击图片查看课程排期

扫码回复【uptime培训】了解课程详情

图片

展览大厅里不同的液体冷却技术

Different liquid cooling technologies on the exhibit floor

根据液体与发热电子元件的相互作用方式，数据中心的液体冷却可分为两大类:芯片直接冷却DtC（译注:冷板式液冷)和浸没式冷却。在DtC配置中，流体不直接与电子元件接触，但在空气冷却配置中，液体冷却剂被泵送到冷板上，以代替电子元件上的散热器。虽然大部分热量都被冷板带走了，但仍然需要风扇来帮助散热，但在某些情况下，风量和风速都很低。有一些设计是通过空气将热量从服务器机箱中交换出来，而另一些设计则需要在机架级或列间安装热交换器，将热量传递到一次冷却回路。

Liquid cooling in datacenters can be divided into two major categories: DtC and immersion, based on how the liquid interacts with the heat-generating electronic components. In a DtC configuration, fluid is not in direct contact with electronic components, but the liquid coolant is pumped to cold plates in place of the heatsinks on electronic components in an air-cooling configuration. While most of the heat is taken away by the cold plates, fans are still needed to help remove heat at the board level, but at a very low volume and velocity in some cases. Some of the designs exchange the heat out of the server chassis through the air while some others require a heat exchanger, either at rack level or row level to transfer the heat to a primary cooling loop.

在浸没式冷却配置中，流体与IT设备直接接触；设备浸在介电流体中。然后产生的热量直接散发到液体中，完全消除了风扇。这通常需要一种不同于传统机架的架构。介电流体是一种不导电但导热的液体，可以安全地与电子元件一起使用。介电流体可以是碳氢化合物(矿物、合成或生物油)或氟碳化合物(完全工程)。DtC和浸没式也可以分为单相和两相，单相即其中液体冷却剂始终保持液体状态，两相即其中液体冷却剂蒸发以传递液体的热量，然后转换回液体状态。提供单相DtC解决方案的公司包括CoolIT、JetCool和Motivair。ZutaCore提供两相DtC解决方案。对于浸没式冷却，Midas提供单相解决方案，而LiquidStack提供两相解决方案。

In an immersion cooling configuration, fluid is in direct contact with the IT equipment; the equipment is immersed in a dielectric fluid. The heat generated is then directly dissipated into the liquid and fans are completely eliminated. This typically requires an architecture different from that of traditional racks. The dielectric fluid is an electrically nonconductive but thermally conductive liquid that is safe to use with electronic components. Dielectric fluid can be hydrocarbon-based (mineral, synthetic or bio-oil) or fluorocarbon-based (fully engineered). Both DtC and immersion also can be categorized as single-phase, in which the liquid coolant always stays in a liquid state, or twophase, in which the liquid coolant evaporates in order to transfer heat out of the liquid and then is converted back to liquid state. Companies providing single-phase DtC options include CoolIT, JetCool and Motivair. ZutaCore provides two-phase DtC. For immersion cooling, Midas offers a single-phase approach while LiquidStack offers a two-phase tank.

图片

浪是否开始往前推了呢？
Has the flood started?

液体冷却并不是什么新鲜事物——人们已经在数据中心中使用了十多年——但它在行业中仍处于起步阶段，因为风冷仍然可以处理数据中心运行中的大部分冷却挑战。然而，TDP和机架密度的加速增长，加上对可持续性的追求，可能会推动液冷在行业中的采用。目前，只有一小部分数据中心采用了液体冷却，但越来越多的人正在了解并安装它。没有一种通用的方法可以满足所有的需求。虽然高密度应用继续增加，其中许多将需要液体冷却，但仍将有大量应用可以通过空气冷却来处理。即使是需要液体冷却的应用，也要根据具体的条件和期望来做出技术决策。在未来的许多年里，液冷可能会很好地在行业中共存，甚至在同一个数据中心内。

Liquid cooling is not new — people have been using it in the datacenter for over a decade — but it is still in its infancy in the industry since air cooling can still handle the majority of cooling challenges in datacenter operations. However, the accelerated increase of TDP and rack density, coupled with the pursuit of sustainability, could boost the adoption of liquid cooling in the industry. Currently, there remains a limited segment of datacenters that have adopted liquid cooling, but more and more people are learning about and installing it. There is no one universal approach to address all requirements. While high-density applications continue to increase, many of which will require liquid cooling, there will remain a large number of applications that can be handled by air cooling. Even for.applications that require liquid cooling, technology decisions will be made based on specific conditions and expectations. Air cooling and liquid cooling may well coexist in the industry, even within the same datacenter, for many years to come.

这一点从会议现场对观众的调查中得到了反映。绝大多数受访者都认为液冷比空气冷却更有效。大约四分之一的观众表示，他们已经在使用液冷，包括单相DtC、两相DtC、单相浸没式和两相浸没式四种情况，另外约20%的观众使用背板热交换器——一种在机架级使用液体的方法，但从技术上讲，这不是一种液冷技术，因为服务器中的热量仍然通过空气传递。大多数听众回答说，他们的公司计划在未来采用液体冷却，其中一半在未来24个月内采用，一半在那之后采用。当被问及哪种类型的液冷技术最有可能被采用时，所有四种类型都被选择了，但更多的受访者计划使用单相DtC和两相DtC，这可以解释为DtC可以使用传统机架架构，修改最少，而浸没式冷却更具破坏性，即使它最终具有更高的冷却能力。

This is reflected by the live survey among the audience at the conference. Respondents overwhelmingly agree that liquid cooling is more efficient than air cooling. About a quarter of the audience said they are already using liquid cooling, with cases in all four categories of single-phase DtC, two-phase DtC, single-phase immersion and two-phase immersion, plus about 20% of the audience using rear door heat exchangers (RDHX) — an approach using liquid at the rack level but is technically not a liquid cooling technology, since the heat in the server is still transferred through the air. Most of the audience answered that their firms plan to adopt liquid cooling in the future, with half adopting in the next 24 months, and half beyond that. When it comes to the question of what type of liquid cooling technology is most likely to be adopted, all four types were selected, but more respondents were planning to use single-phase DtC and two-phase DtC, which can be explained by the fact that DtC can use legacy rack architecture with minimal modifications, while immersion cooling is more disruptive even though it ultimately has a higher cooling capability.

无论是DtC还是浸没式，都可能需要在空余空间或设施级别对服务器、机架和基础设施进行一定程度的修改。虽然数据中心行业已经围绕空气冷却轻松地建立了其标准和系统之间的接口，但液体冷却的出现带来了重新定义和重建标准的挑战。这需要从芯片供应商、服务器和基础设施、数据中心运营商、设计和建造公司到监管机构的全面努力。

Whether it is DtC or immersion, a certain level of modifications to the server, rack and infrastructure at the white-space or facility level may be required. While the datacenter industry has comfortably established its standards and interfaces between systems and players around air cooling for decades, the emergence of liquid cooling brings the challenge of redefining and reestablishing standards. This will require a comprehensive effort, from suppliers of chips, servers and infrastructure, datacenter operators and design and build firms to regulatory bodies.