当前位置：首页 > news >正文

液冷与电力能否在同一机架中安全共存？

news 2026/8/13 9:49:58

作者：Ryan Jarvis, Vertiv

Vertiv 的全球服务副总裁兼负责人

vice president and head of global services at Vertiv

人工智能已经重新绘制了数据中心基础设施的热力与电力分布图。但当液体与电力混合时，安全并非理所当然；它是被设计出来的。

AI has redrawn the thermal and electrical map of data center infrastructure. But when liquid and power mix, safety is not assured; it’s engineered

AI加速已经超越了摩尔定律，突破渐进式增长，在更小的空间内实现了功率、热量和密度的指数级提升。

现在机架密度已超过140千瓦（kW），现代AI处理器的热设计功率（TDP）突破1000瓦（W）。这些需求正迫使高性能计算（HPC）设计发生转变，其中液冷和中压电力必须共存于同一机架。

在机架层面集成液体和电力并非挑战。真正的挑战是在多个机架和站点间安全地扩展集成。随着AI密度攀升，风险也在增长，故障的后果同样如此。大规模下的可靠性取决于团队如何管理四个关键领域：电力安全、流体质量、防泄漏以及运营准备。

AI acceleration has outpaced Moore’s Law, pushing past incremental gains to deliver exponential increases in power, heat, and density, all within a smaller footprint.

Rack densities now exceed 140 kilowatts (kW), and modern AI processors blast past 1000 watts (W) thermal design power (TDP). These demands are forcing a shift in high-performance computing (HPC) design where liquid cooling and medium-voltage power must share the same rack.

Integrating liquid and power at the rack isn’t the challenge. Scaling integration safely across many racks and sites is. As AI density climbs, risks grow. So do the consequences of failure. Reliability at scale depends on how well teams manage four critical areas: power safety, fluid quality, leak prevention, and operational readiness.

HPC机架通常在中压水平下运行，增加了维护期间的电弧闪光风险。部署的机架越多，出错的可能性就越大。安全操作依赖于严格的流程、经过培训的人员以及IT与设施之间的紧密协调。

HPC racks often operate at medium-voltage levels, increasing arc flash risks during maintenance. The more racks you deploy, the more chances for error. Safe operation depends on strict procedures, trained staff, and close coordination between IT and facilities.

更高的热量需要更好的流体管理

高密度冷却系统中的流体质量是一项持续的责任，而非一次性的检查。即便是颗粒或微生物滋生等微量污染物，也会随着时间的推移降低热传导或冷却能力，并损坏冷板。冷却回路中的气泡会产生起泡效应，扰乱冷却液循环，降低效率，并给水泵带来负担。

这些都是真实且代价高昂的问题，若不进行主动管理，它们可能会悄然加剧并造成严重打击。可靠性将取决于整个冷却回路的安装与日常精准维护。

我们不再仅仅是在机架附近添加液冷。我们正在进行在运行中的芯片上方进行冷却。一个针孔大小的泄漏、一个松动的接头，或最微小的热膨胀计算失误，都可能导致整个机架停机、处理任务丢失以及昂贵的宕机时间。在这样的密度下，微小的故障会带来不成比例的后果。

More heat requires better fluid management

Fluid quality in high-density cooling systems is an ongoing responsibility, not a one-time check. Even trace contaminants like particles or microbial growth can reduce thermal heat transfer or capacity and damage cold plates over time. Air bubbles in the cooling loop create a frothing effect, disrupting coolant circulation, reducing efficiency, and straining pumps.

These are real, costly issues that can grow silently and hit hard without proactive management. Reliability will depend on installation and daily precision across the entire cooling loop.

We're no longer just adding liquid cooling near the racks. We're cooling on top of live chips. A pinhole leak, a loose fitting, or the slightest thermal expansion miscalculation can lead to complete rack shutdowns, lost processing, and expensive downtime. At these densities, small failures carry outsized consequences.

更高的复杂性需要新技能

传统的数据中心团队技能娴熟，但高性能计算（HPC）数据中心中液冷的兴起带来了一些维护任务，这些任务所需的专业知识在传统IT部署中并不常见。诸如压力管理、液体取样、泄漏检测和冷却液更换等任务需要专门的培训和流程。

学习曲线是显著的：团队必须在大规模适应新应用的同时，保持可靠性和正常运行时间。弥合这一准备度差距，对于在实时环境中安全地扩展液冷至关重要。

More complexity demands new skills

Conventional data center teams are highly skilled, but the rise of liquid cooling in HPC data centers has introduced maintenance tasks that require expertise not commonly found in traditional IT deployments. Tasks such as pressure management, fluid sampling, leak detection, and coolant replacement necessitate specialized training and processes.

The learning curve is significant: teams must adapt to new applications on a large scale while maintaining reliability and uptime. Bridging this readiness gap is essential to safely scaling liquid cooling across live environments.

调试：安全的第一道防线

每次安全的部署都始于通电之前。在施工期间，液冷维护人员遵循一条铁律：在通电之前，机械系统必须经过彻底评估。

调试在受控条件下将高密度冷却模型推至极限，在真实工作负载接触它们之前暴露故障。系统以故障模式运行，协同进行压力测试，并且理应暴露缺陷，因为那些时刻能告诉我们设计是否可以被强化并再次测试。

每根管道都经过压力测试，每个接头都经过检查，冷却回路被冲洗直至流体达到精确的纯度规范。当出现故障时，不会仅作记录就置之不理。它会得到修复并经签字确认。团队进行实时调整，修复薄弱点，并重新测试，直至系统在整个运行压力下的表现符合预期。

Commissioning: The first line of safety

Every safe deployment begins before the power switches on. During construction, liquid cooling servicing staff follow an ironclad rule: mechanical systems must undergo thorough evaluation before energizing power.

Commissioning pushes high-density cooling models to their limits under controlled conditions, revealing faults before real workloads touch them. Systems are run in failure modes, stress-tested in tandem, and expected to expose flaws because those are the moments that tell us whether the design can be hardened and tested again.

Every pipe is pressure-tested, every joint inspected, and cooling loops flushed until the fluid meets exact purity specifications. It's not logged and left when something fails. It’s fixed and signed off. Teams make real-time adjustments, repair weak points, and retest until the system performs as intended under whole operational stress.

工程师从一开始就与客户坐在一起，制定调试计划、故障处理方案和长期维护策略。这种合作关系并不会在系统上线时终止——瑞安·贾维斯，维谛技术

Engineers sit with clients on the get-go, mapping out commissioning plans, failure protocols, and long-term maintenance. That partnership doesn’t end at go-liveRyan Jarvis, Vertiv

维修团队只有在评估了液体系统的完整完整性后，才会开始电气测试。这种严格的顺序——先机械调试，后电气调试——可防止灾难性的重叠，例如在隐蔽泄漏未被发现时测试 700 伏（V）系统。如果合作方并行调试电力和冷却系统，容错余地往往会相对狭窄。

Servicing teams begin electrical testing only after assessing the complete integrity of the liquid systems. This strict sequencing, where mechanical commissioning goes first before electrical, prevents catastrophic overlaps, like testing a 700-volt (V) system while a hidden leak escapes detection. If partners commission power and cooling in parallel, the margin for error tends to be relatively narrow.

运营成熟度：从设计到全生命周期管理

一旦电力和液冷系统投入运行，可靠性所依赖的不仅仅是安装；它取决于内部客户工程师（CE）在高压环境下设计的高密度冷却系统的响应能力。

这始于内置的安全保障。液体与电力之间的每个接口都采用分层防护设计：隔离区、受监控的泄漏托盘和自动关闭协议。这些并非事后补救。它们定义了系统如何缓解并消除固有风险。

工程师从一开始就与客户坐在一起，制定调试计划、故障协议和长期维护方案。这种合作并不会在系统上线时结束。它会通过日常支持持续下去：液体采样、操作员培训、系统审查，以及随着计算需求演变而进行的性能调优。

先进的预防性维护使针对性维护成为可能，团队在常规服务清单之外，重点关注高影响项目。此外，实时监控可检测潜在异常。这使团队和咨询工程师能够立即介入，防止故障并提升可靠性。

Operational maturity: From design to lifecycle management

Once power and liquid cooling systems are operational, reliability hinges on more than installation; it depends on how well internal customer engineers (CE) have designed high-density cooling systems to respond under pressure.

That starts with built-in safeguards. Every interface between liquid and power is engineered with layered protection: separation zones, monitored leak trays, and automated shutdown protocols. These aren’t afterthoughts. They define how the system mitigates and eliminates inherent risks.

Engineers sit with clients on the get-go, mapping out commissioning plans, failure protocols, and long-term maintenance. That partnership doesn’t end at go-live. It continues through daily support: fluid sampling, operator training, system reviews, and performance tuning as computing demands evolve.

Advanced preventive maintenance enables targeted maintenance where teams focus on high-impact items on top of the usual servicing checklist. Additionally, real-time monitoring detects potential anomalies. This enables teams and consulting engineers to intervene immediately, preventing failures and enhancing reliability.

为安全而设计，安心运营

存在诸如液体污染、泄漏以及电弧闪光等风险。但这些并不是回避液冷的理由，而是应当严谨对待该解决方案的理由。通过恰当的设计、调试规程以及运行规范，液冷既安全，又是在大规模场景下保持性能与正常运行时间的理想选择。

Design for safety, operate with confidence

There are risks, from fluid contamination and leaks to arc flashes. But these aren’t reasons to avoid liquid cooling, rather to approach the solution rigorously. With the right design, commissioning protocols, and operational discipline, liquid cooling is both safe and ideal for maintaining performance and uptime at scale.

维谛技术Vertiv

维谛（Vertiv）是全球领先的关键数字基础设施解决方案提供商，广泛服务于数据中心、通信网络以及各类商业与工业场景。

作为关键数字基础设施领域的全球领导者，维谛技术拥有业界完备的产品组合。