---
格式版本: 2
标题: "Liquid Cooling for AI: Advanced Architecture Explained | Schneider Electric"
原文链接: "https://www.youtube.com/watch?v=PTsOuxwE9Oo"
发布日期: "2026-04-10"
发现时间: "2026-04-10T00:00:00+08:00"
入库时间: "2026-05-28T02:39:37+08:00"
来源平台: "历史资料迁移"
搜索渠道: "legacy_migration"
搜索词: "url-index.json 迁移"
匹配关键词:
  - "AI Rack"
  - "液冷"
  - "散热子系统"
  - "url-index.json 迁移"
  - "待定数据"
  - "关键字"
  - "YouTube"
  - "10.md"
相关厂家:
  - "NVIDIA"
相关专家:
  []
内容类型: "视频"
抓取工具: "历史采集"
清洗工具: "历史资料原文保留 + LLM 正文裁剪"
关联判断: "待人工确认"
关联置信度: 0
关联理由: "历史资料迁移，未重新调用模型判定"
AI优质: "是"
AI打分: 89
AI分档: "高置信优质"
AI质检状态: "通过"
AI打分理由: "施耐德与联想专家深度探讨AI数据中心液冷架构，涵盖冷板热阻、CDU分级部署、冷却液选型及单相/两相技术路线，技术细节扎实，含实际部署经验与TCO分析，高度契合超节点散热与基础设施主题。"
AI质检模型: "qwen3.6-plus"
AI质检时间: "2026-06-03T00:34:45+08:00"
AI主题相关性: 18
AI来源权威性: 14
AI新颖性: 16
AI技术细节: 19
AI商业部署信号: 13
AI完整性: 9
采集批次: "2026年5月28日2点39分33秒"
采集批次ID: "legacy-migration-20260528023933"
去重键: "https://www.youtube.com/watch?v=PTsOuxwE9Oo"
---

## 视频描述

本次网络研讨会将探讨适用于 AI 级数据中心的先进液体冷却架构，包括 Delta T 方程如何影响容量规划和性能表现。了解机械系统为何会带来独特的设计挑战，流量/压力/温度限制如何影响高密度计算，以及来自实际部署中的常见陷阱。

- 液体冷却中的机械挑战与电气挑战
- 流量、压力和温度限制
- Delta T 在系统性能中的作用
- 现场部署中的经验教训

## 字幕全文

**[00:00:00]** 早上好，下午好，晚上好。我是施耐德电气的 Rob Bunger，非常高兴能

**[00:00:07]** 主持本次网络研讨会，我们邀请了一些在液冷领域真正有实战经验的专家。

**[00:00:15]** 目前许多人正在部署液冷，但能够亲自动手并了解设计与部署各项复杂细节的人

**[00:00:23]** 并不多，而我们正好有两位这样的专家。

**[00:00:29]** 首先我想介绍 Motivair 的 Greg，还有 Lenovo 的 Vinod。

**[00:00:36]** 那么 Greg，请你自我介绍一下吧。好的，很高兴来到这里。简单介绍一下，我是 Greg Alexander，

**[00:00:44]** 施耐德电气旗下 Motivair 的高级热工工程师。我在高性能计算领域拥有丰富的

**[00:00:50]** 经历。职业生涯始于 Cray，后来进入 HPE，主要专注于

**[00:00:57]** 高密度、百分百液冷的服务器，包括冷板设计、导热板设计，以及完整的

**[00:01:06]** 热系统方案，研究流体管路和布局。

**[00:01:11]** 现在在 Motivair，我是往热链的更高一层发展，聚焦于不仅仅是 GPU，

**[00:01:20]** 更多关注 CDUs（冷却液分配单元）以及我们如何将流体真正部署到 IT 机架中。

**[00:01:29]** 很棒，谢谢 Greg。感谢你的加入。Vinod，请你介绍一下自己。

**[00:01:34]** 好的，我叫 Vinod Kamath，在北卡罗来纳州的 Lenovo 工作。

**[00:01:41]** 加入 Lenovo 之前，我的团队属于 IBM，我们负责高密度

**[00:01:48]** 高性能液冷和风冷的 X86 解决方案。

**[00:01:53]** 我是该领域产品的架构师，随着我们转型至

**[00:01:59]** Lenovo，我们带来了为高密度

**[00:02:07]** 温水液冷方案开发的技术，实现了能效计算，

**[00:02:13]** 服务于混合性能计算领域。现在我领导 Lenovo 的高性能与 AI

**[00:02:21]** 系统设计架构，并部署几乎百分百液体捕获冷却解决方案。

**[00:02:29]** 好，非常棒。本次讨论会涉及一些技术细节，但不会太复杂。

**[00:02:37]** 我想先从简单的地方开始，Greg，先问你一个问题。

**[00:02:42]** 好的。你觉得为什么水比空气更好？从根本上讲？我们探讨物理学，从流体的属性分析

**[00:02:53]** 空气与液体之间，如果我们参考质量流量公式，

**[00:02:58]** 即 Q-dot = m-dot × CP × ΔT。Q-dot 是你的

![关键帧 00:02:58](./assets/20260424_PTsOuxwE9Oo_178s.jpg)

> 视觉分析：标题：Liquid Cooling Architecture: Advanced Design Principles for AI-Class Data Centers

**[00:03:04]** 功率，m-dot 是质量流量，CP 是热容量系数，

**[00:03:09]** 或称为比热容。ΔT当然是流体的

**[00:03:15]** 起始与终点温度差。所以，空气和液体相比，液体，比如

**[00:03:22]** 去离子水或 PG25，这些都是行业常用的冷却液。

**[00:03:27]** 液体密度更大，所以液冷的质量流量远高于风冷，水的比热容也更好。

**[00:03:36]** 我喜欢把比热容比作桶——你能装多少焦耳的热量。

**[00:03:42]** 桶越大，就能带走更多热量。

**[00:03:48]** 举个例子，比如一个 500 瓦的 GPU，就是你的 Q，这是个固定值。

**[00:03:54]** 如果用液冷或风冷，风冷的情况是

**[00:03:59]** 质量流量较低、比热容较低，ΔT 就会非常大。

**[00:04:05]** 但引入液体后，你获得更高的质量流量与比热容，

**[00:04:10]** ΔT 可以更低，因此整个系统效率就提升了。

**[00:04:16]** 很好。Vinod，你们在很多年前决定

**[00:04:22]** 把风冷服务器换成液冷时，你们的决策流程是什么？

**[00:04:31]** Rob，这个问题很好。过去我们不用面对

**[00:04:37]** 如 Greg 所说的那些 GPU。十年前左右，

**[00:04:43]** 但我们开始在市场上，特别是

**[00:04:48]** 欧洲，看到能源价格很高。在德国一度电是 0.20 欧元。

**[00:04:57]** 那是 2010 年。如果你考虑部署和运营高密度

**[00:05:07]** 高性能计算集群的总成本，能源价格就很关键。

**[00:05:13]** 其次，若有当地规定要求更高效的能源管理，

**[00:05:19]** 减少水消费等。正如 Greg 所说，液冷的效率驱动着

**[00:05:30]** 用最少能源实现冷却和最大化计算输出。

**[00:05:38]** 我们当时首次部署的，是一个 10,000 节点的 X86 集群。

**[00:05:48]** CPU 只有温和的热设计电力，但在整个运行周期内，

**[00:05:54]** 客户节省了大量成本。这也是我们在 IBM 和 X86 开始探索的起点。

**[00:06:02]** 随着时间发展，我们在密集的 X86 以及 GPU

**[00:06:08]** 领域提供极致效率和性能，正如 Greg 所述，

**[00:06:16]** 液体能在更小体积内捕获更多热量，相比空气。

**[00:06:23]** 很有趣的是，最早的客户其实非常关注效率，

**[00:06:29]** 这一直是液冷的优势。现在我们又得到了双重好处：

**[00:06:37]** 不仅能冷却市场最热的芯片，高效也成了副产品，你还可以在

**[00:06:47]** 相同的功率包络内部署更多算力，因效率提升。

**[00:06:53]** 我想深入探讨一下冷板上到底发生了什么，

**[00:07:01]** 听说过 T-case、热阻等专业术语，

**[00:07:07]** 当你考虑液冷性能时，这些概念如何协同作用？

**[00:07:14]** Greg，我们先从你开始。好的。其实我们可以非常深入地

**[00:07:21]** 探讨，但这里还是简单点吧。就像，从

**[00:07:26]** 热力学定律出发，就是我们之前提到的公式，但其中也有很多细节问题，

**[00:07:33]** 定义各种术语。你提到的热阻，是一个大框，能

**[00:07:40]** 不是说要简化系统，而是让它更容易理解。其实你看典型

![关键帧 00:07:40](./assets/20260424_PTsOuxwE9Oo_460s.jpg)

> 视觉分析：标题：Liquid Cooling Architecture: Advanced Design Principles for AI-Class Data Centers

**[00:07:50]** 电阻公式 V=IR，两者完全可以类比。

**[00:07:55]** 热阻对应电阻（欧姆）。在电学类比中，电流更像 ΔT，

**[00:08:08]** 但 I 是你的 Q，是你的功率流，ΔT 就像电压，

**[00:08:16]** 差异点就是你传递功率的两个端点。所以通过这种方式，公式就变成了热阻

**[00:08:26]** 单位通常是每瓦摄氏度（°C/W）。

**[00:08:31]** 这样只有一个数字，可以更直观体现

**[00:08:36]** 冷却效率。举个例子，还是比较空气与水，普通

**[00:08:43]** 顶级风冷散热器用于 GPU 的热阻大约在

**[00:08:51]** 而高性能液冷冷板通常优一个数量级，达到 0.02 °C/W，非常典型，

**[00:08:56]** 有些效率不那么高，有些则能达 0.01 或甚至更低，具体取决于几何结构，

**[00:09:04]** 但这充分说明液冷在物理层面优于风冷。

**[00:09:10]** 如果要深入讨论，还可以找电容的类比等其他点。

**[00:09:17]** 比如电容器等等，你能在类比中找到对应的热学等价物。

**[00:09:22]** 很棒，我很喜欢这种电学类比的解释。

**[00:09:28]** 这确实帮助我理解。完全正确。

**[00:09:34]** 最近行业讨论很多，Vinod，这个问题问你。现在有 PG25，

**[00:09:43]** 行业好像不敢说完全统一，但经常听到

**[00:09:49]** 大规模液冷系统部署时都在用PG25，当然还有去离子水（DI），Lenovo也有丰富经验，

**[00:09:59]** 能否谈谈怎么做选择？好的，正如你所说，单相液冷目前

**[00:10:09]** 有两种常用冷却液：DI水，就是Lenovo在

**[00:10:16]** 从2010年起一直采用，运行可靠，效率

**[00:10:24]** 略高。此外某些地区，地缘因素也促使客户选择更容易监管的液体。

**[00:10:34]** 不同国家会更偏好环境友好的流体。

**[00:10:40]** 但这要求客户、解决方案提供商必须细致维护

**[00:10:51]** 化学配方。DI水是生物敏感液体，

**[00:10:59]** 有细菌滋生风险——毕竟它确实就是水，所以必须每季度维护化学成分，

**[00:11:09]** 只要你控制好配方管理，

**[00:11:14]** 我们历史上证明它可靠。至于 PG25，近年来行业里

**[00:11:23]** 液冷方案呈现爆发式增长，现在不仅是实验室或科研

**[00:11:32]** 用于计算，更多的企业环境也开始用液冷，

**[00:11:40]** 但客户往往没相关专业知识，供应商在生态链里也多样，

**[00:11:47]** 独立控制液体化学配方变得极其困难。

**[00:11:53]** PG25 在这方面帮了大忙，因为 25% 的乙二醇含量使其

**[00:12:04]** 基本生物惰性。还是要小心，但生物滋生风险很低。

**[00:12:10]** 控制好后，剩下就是化学配方维护。即使如此，PG25也要有维护周期

**[00:12:20]** 做好维护后，才能真正享受

**[00:12:28]** 比所有风冷系统更高的可靠性。

**[00:12:34]** 管理得当的水和合理管理的 PG25都能在任何规模、任何环境

**[00:12:40]** 带来客户想要的效率。

**[00:12:48]** 很有意思。那么，我理解你刚才的意思是？

**[00:12:54]** 即便在风冷环境���，像有灰尘等各种问题，

**[00:13:03]** 从历史看，液冷系统的可靠性与性能都优于风冷。

**[00:13:09]** 是的。我们确实有这方面的数据表现。（遗憾的是还未公布内部白皮书）

**[00:13:15]** Lenovo内部数据建议如下，这也呼应了我们前面讨论的热阻，

**[00:13:21]** 以及液冷的固有优势。

**[00:13:28]** 液冷能为硅芯片带来更窄的温度区间，

**[00:13:36]** 此外流速与温度都能几乎保持

**[00:13:42]** 平均值只波动几个百分点。

**[00:13:47]** 流速浮动只有正负 2%（针对机架），

**[00:13:54]** 温度被 PID 精确控制，波动不超过一摄氏度，

**[00:13:59]** 风冷系统则不可能做到。风冷扇区会随着设备温度动态调整风量与芯片温度。

**[00:14:07]** 我们发现芯片在液冷下能稳定运行在狭窄温度区间，

**[00:14:17]** 提高芯片可靠性。

**[00:14:23]** 当然仍可能发生故障，但故障不是温度造成的，更可能是固件

**[00:14:33]** 或某些芯片封装问题，所以导致异常。

**[00:14:40]** 很有意思。回到水冷化学管理，

**[00:14:48]** 这个针对 Greg 和你。你提到

**[00:14:54]** 对 DI 水一般每季度维护一次？

**[00:14:59]** 是的。测试等都在每季度进行。那 PG25呢？维护周期是否一样？

**[00:15:06]** PG25 维护方面，以下是我们的实践：

**[00:15:13]** 水的话就是季度点检，操作简单但必须进行。PG25的话，

**[00:15:19]** 很多客户会根据情况，

**[00:15:31]** 维护周期可适当放宽，也取决于最终用户，

**[00:15:39]** 也许你可以把季度维护

**[00:15:45]** 放宽到每半年一次。个人观点是，这些服务器的成本如此昂贵，

**[00:15:56]** 所以还是应该认真管理化学配方，

**[00:16:05]** 每半年花点小钱做校验。

**[00:16:10]** 不是难事，很容易操作，基本上很微不足道。

**[00:16:15]** 建立好标准流程后就没问题。你提到成本，很重要——和冷却系统本身相比，这点费用

**[00:16:21]** 根本就是九牛一毛。我的观点也类似，对于DI水需要更严格

**[00:16:30]** 把控取样和监测，这也正是行业更偏向PG的原因。大家觉得

**[00:16:36]** 其实PG25并不意味着永远不会出现微生物滋生，你还是要管理，

**[00:16:44]** 而且在系统投产时要落实好。正如你所说，管理依旧不能放松，

**[00:16:51]** 尤其服务器供应商要保证产品质量，

**[00:16:58]** 我们服务器有保修，所以我们必须掌握化学配方状况，决不能掉以轻心。

**[00:17:08]** 无论任何现场部署，都严格管控化学配方。正如我听说的，

**[00:17:16]** 如果失控，将会是噩梦，恢复成本极高。

**[00:17:21]** 虽然不想渲染恐怖气氛，但面对新一代设计，

**[00:17:29]** 冷板间距更细，需要用微米级过滤器，鳍片密度越来越高，

**[00:17:38]** 结构会变得更加复杂，三维几何越来越细腻，

**[00:17:44]** 内部出现任何异物都难以清理。

**[00:17:53]** 所以一定要小心。不过事实上，只要管理得当，

**[00:18:00]** 有客户已成功部署水冷系统并稳定运行六至八年。

**[00:18:06]** 所以事实上这不是新技术。

**[00:18:11]** 很好的观点。工程学没问题。现在话题稍微转换下。

**[00:18:17]** 我们谈到了服务器层的情况，还有水质。现在要问的是，

**[00:18:28]** 服务器？这就涉及到 CDU。Greg，CDU的作用是？

**[00:18:33]** 它主要有哪些功能？为什么存在？好的。从总体来说，

**[00:18:39]** 它包含几大核心部件：换热器，泵，过滤器，通常还要配合液体化学管理

**[00:18:47]** 阀门、压力计、流量计，这些组成了系统的“防护墙”。

**[00:18:53]** 通常有两套循环系统，业界已经普遍采用TCS和FWS术语。

**[00:19:00]** TCS即技术水系统或技术水冷系统，

**[00:19:06]** 就是我们前面讨论的直接供冷到芯片的流体。

**[00:19:11]** FWS则是设施水系统。设施水系统由来已久，

**[00:19:17]** 传统风冷服务器往往向室内送出热风，热量必须消散，

**[00:19:23]** 根据热力学原理。机房侧墙设有CRAC、CRAH设备，将空气热量转移到

**[00:19:31]** 屋顶的冷水机、冷却器、或水侧节能器等等终端设备。

**[00:19:37]** CDU就在这个生态里，成为技术流体与设施水的隔离屏障。

**[00:19:48]** 与历史上设施水不同，设施水系统有铁管等杂质，我们不希望这些

**[00:19:54]** 接近IT设备。所以说，归根到底，

**[00:19:59]** 很形象的描述。我一直把它比作电气里的变压器，

**[00:20:07]** 相当于把两路流体分隔开了。

**[00:20:13]** 对，没错。CDU能控制质量、温度、压力等参数。好的。

**[00:20:18]** 很棒。CDU类型很多，从机架CDU（只服务单列IT设备）

**[00:20:27]** 到超大型CDU及TCS大循环。

**[00:20:34]** 能否两位谈谈部署不同CDU类型的权衡，Vinod先来分享。

**[00:20:43]** 你遇到什么情况？高性能计算部署时，

**[00:20:50]** 通常采用机房级解决方案，因为是单一物理空间。客户通常买10到40个机架，

**[00:21:01]** 极端甚至100个机架。如果做TCO分析，客户能管控好

**[00:21:08]** 化学配方的话，适合部署机房级0.5-1.5兆瓦CDU

**[00:21:18]** 能按拓扑灵活部署，Greg可以展开说N+N、N+1等。

**[00:21:24]** 但现如今快速AI算力消费和企业化需求增长很快，

**[00:21:33]** 在这种模式下，你并不了解设施水回路，也没有时间改造或重新设计，

**[00:21:42]** 所以可以管理机架级CDU，直接连接CDU，同时

**[00:21:53]** 也可以机架集成CDU，单列化学分隔管理。简而言之两者都存在，

**[00:22:05]** 新趋势更偏向小型、机架级CDU，客户买几台机架就用。

**[00:22:11]** 趋势是客户只买几个机架。没错。我想用“rack and stack”这个词，

**[00:22:19]** 不是我原创，但人们很喜欢它——即每次只部署一台机架，

**[00:22:25]** 所有液冷集成在机架里，接入设施系统就能直接用。只需搬到机房就能立刻运行。

**[00:22:33]** 如果用地面大型CDU，服务多台机架、

**[00:22:41]** 承载能力更高，需要同步部署大规模设备。

**[00:22:46]** 我喜欢这种快速部署方式。

**[00:22:51]** 从权衡角度说，如果每个机架都配CDU，

**[00:23:01]** 单台CDU成本应该高于中心化系统，同时周期性

**[00:23:08]** 化学测试维护也会增加，长远来看，

**[00:23:13]** 机架CDU总成本可能更贵。

**[00:23:20]** 我来谈谈Lenovo的视角。其实这取决于具体情况，

**[00:23:30]** 如果是大规模N+N或N+1机房，

**[00:23:35]** 现场有运维工程师驻守，

**[00:23:41]** 部署规模足够大，数据中心级CDU（如1.5兆瓦）往往更容易管理，

**[00:23:51]** 控制成本也更优，毕竟现场专人实时监控，

**[00:24:00]** 日常维护，有专业人员。

**[00:24:05]** 即使部署大，但如果没有

**[00:24:12]** local expertise, then the question that
you might want to ask is, do you have

**[00:24:17]** a service contract with a CDU vendor? Where, right. You know, it's, they are sort
of still managing that at the

**[00:24:24]** node level or the rack level. Right? So it's in, in rack CDU integrated, but
managed by the, you know, the provider

**[00:24:32]** of, of liquid services and you know,
that cost may be something, that,

**[00:24:38]** you know, Motivair can comment on. It may be higher, but
it need not be higher. Yeah.

**[00:24:44]** Yeah. And I think total cost of ownership is
a, is a good phrase to use here, because I, I can't get into the business of it.

**[00:24:49]** I'm more the, you know, the physics
guy, the nerd in the back closet. But, when, when it comes to, you
know, yeah, looking at total cost of

**[00:24:56]** ownership, that the, there could be some
things to consider as far as service level and what's easier to deploy.

**[00:25:02]** I think from a fundamental physics
perspective, maybe an analogy, I could use this on the air cooling side.

**[00:25:07]** It's just, it's really just an economy of
scale where you're gonna get potentially more efficient gains from a, you know,
1.5 megawatt CDUs we've been using.

**[00:25:16]** if you compare that to like an air quote
system, imagine like 140 millimeter fans running on all those servers versus two
or three large blowers on a fan wall.

**[00:25:25]** You know, it, it's pretty clear,
obviously which one's gonna be more efficient at moving, you know,
in bulk, a large amount of air.

**[00:25:31]** So it's not a direct comparison, but
it's, it's in an economies of scale. Yeah.

**[00:25:39]** When we talk about, heat transfer, getting
back to the physics, there's still, you

**[00:25:47]** know, a lot of different liquid cooling
technologies out there and, you know, two phase direct to chip is something that,
you know, you, you hear talked about a

**[00:25:55]** lot in the market, and I'd like to see
if you can, you know, address what that is and, and why it's good and, and, you
know, any, any challenges to deploying it.

**[00:26:06]** And, and Greg, I'll, I'll
start with you on, on that. Sure, yeah. Yeah. It's certainly an exciting field.

**[00:26:12]** I mean, obviously refrigeration has been
used in the, you know, air conditioning cycles for, for what, a hundred years now?

**[00:26:18]** Maybe not that long, but we've been
aware of, you know, the car cycle and all these different technologies that
use ways to, vapor compression allow

**[00:26:26]** for, you know, refrigerant to cool
things at higher temperatures than what they're trying to cool too, right?

**[00:26:32]** And so if we get back to the physics
part of it, what you're doing is you're, you're boiling a fluid, right?

**[00:26:38]** And so what you take advantage of there
is the latent heat of vaporization. So essentially you're gonna be
providing liquid refrigerant to the

**[00:26:47]** cold plate, and then it boils directly
on the heat source of the cold plate. And then li's as typically
a, a vapor mixture.

**[00:26:54]** It's not gonna be leaving as a
hundred percent vapor, for, for some safety reasons maybe we can get into. But, the idea is it's a lot more
localized heat flux that you can

**[00:27:02]** get out of a very small package. Whereas, so if you compare that to
like a single phase system, you're

**[00:27:08]** kind of in a, in a linear path, right? Where if you need more cooling, typically
you're gonna need more flow because that's

**[00:27:15]** your, that's your, now that you can turn. So you, you typically will need a
lot less mass flow because a lot of

**[00:27:21]** that heat is not necessarily just
going into heating up the fluid, it's going into boiling the fluid.

**[00:27:26]** So obviously your velocities
might be a little bit higher on the vapor side, right? So there's gonna be some pressure
drop considerations there.

**[00:27:33]** But the idea behind two-phase, you know,
being superior or not superior, that that's, you know, up for debate, right?

**[00:27:39]** But the idea is that at the local
GPU, you're able to get a much higher influx and that that's
why you would end up using that.

**[00:27:46]** Okay. Another advantage to that,
that right is do, do you have a single boiling temperature.

**[00:27:52]** So a lot of GPUs are getting some
really weird mixed topologies these days where you might have hotspots
in spots that aren't typical to, you

**[00:27:59]** know, CPUs from 10, 20 years ago. And so those hotspots that are managed,
because the entire services is just

**[00:28:04]** at the boiling temperature, you don't
have to worry about, you know, keeping in different spots within the package.

**[00:28:10]** Interesting. Vinod, is this something that, you
know, whatever you can share if you guys

**[00:28:16]** researched or your thoughts on that? Yeah. So the way I'd like to answer this
question is the following, right?

**[00:28:22]** So at the end of the day, as a server
solution provider, what we bring

**[00:28:29]** to the, sort of end customers is
technologies that are necessary to keep the silicon working reliably
and, you know, efficiently, right?

**[00:28:37]** And, and have, so of an
expectation that the invested technology has some life, right?

**[00:28:44]** It's available and then it's, managed
to, the customer expectations.

**[00:28:51]** Today, what we are seeing is on
from the data available to us that

**[00:28:57]** all the next, few years, right? We can still, keep using single
phase liquid cooling and deliver the

**[00:29:05]** cooling expectations of accelerators
that are pushing, you know, more than, kilowatt and a half, right?

**[00:29:13]** I mean, there, there's data available
in our labs that document that behavior.

**[00:29:18]** So for us, the perspective that
we have in this space is use the

**[00:29:26]** technology that works and research in
the technology for the future, right?

**[00:29:32]** Because it is to the
point that Greg made. I mean, you may run into a scenario
where there are some local temperature

**[00:29:39]** gradients that need some, you know,
fairly aggressive heat transfer

**[00:29:45]** because the power densities are high. Maybe not the power is high, or
maybe the rate at which the power is

**[00:29:52]** increasing is not, natively aligned
to the most efficient heat transfer.

**[00:29:58]** Right? So we are researching these technologies. We are planning for the future, but, it's
not something that we absolutely need to

**[00:30:09]** deliver the silicon performance in 2026
and 2027, but to the point that has been

**[00:30:15]** made, these technologies exist today. I mean, you will see, of solution
providers delivering two phase cooling,

**[00:30:23]** and there's a place for that based on,
you know, their end customer expectations of, of, efficiency or reliability that
they want. But it is not necessary,

**[00:30:34]** for 2026 products, or probably even
2027 products as best I can tell you.

**[00:30:40]** Perfect. Yeah. And, and I think an important point
to bring up too is, is the, you know, deployment needs and there is
really no free lunch here, right?

**[00:30:48]** Like, it, it's all well and good
to look at, you know, just at the cold plate and say, wow, this
is so much more efficient here.

**[00:30:54]** But then whatcha gonna do with that
vapor, you know, in, in a typical, you know, single phase liquid cooled
system, you just have, you have some

**[00:30:59]** simple hoses, maybe even just some
quick disconnects, that can, you know, be able to swap out servers.

**[00:31:06]** But once you put high pressure
vapor into the mix, hoses are less,

**[00:31:11]** reliable and you need to have a
lot more pressure containment. There's, you know, safety concerns
for, let's say you over boil and

**[00:31:18]** you, you know, I mentioned 70% of a
vapor or vapor liquid mixture, right? If you get into a super heated vapor,
then your pressure's really gonna

**[00:31:26]** start building up 'cause there's
nowhere for that heat to go other than just higher temperature vapor. Higher temperature vapor
gets very high pressure.

**[00:31:34]** So there, there are things that come
with a two phase system that are not

**[00:31:39]** inherently better than, than single phase. I think single phase, like you said,
for, for today is perfectly reasonable.

**[00:31:45]** And, and knowing when that switch over
point is, you know, I'd, I'd like to go get a lot lottery ticket if I knew.

**[00:31:51]** Fantastic. That's great. Thank you. Let's talk supply water

**[00:32:01]** temperature to the rack, to the chip. Yeah. You know, I think, it was interesting
and I think it, CES at the beginning of

**[00:32:11]** this year, you know, kind of in a public
forum, you know, Jensen from NVIDIA had said, Hey, you know, 45 C water
and what's that mean for the facility?

**[00:32:20]** And, you know, let's talk about, you
know, what do you guys recommend or what do you see and you know, the,
the supply water temperature role

**[00:32:29]** and Vinod, I'll start with you. Yeah. So you bring up a good point, right? So in, in January there's
a discussion of warm water.

**[00:32:36]** Maybe it is a weather
that made that up, right? So it's freezing out here in North
Carolina and the Northeastern.

**[00:32:43]** So, I mean, jokes aside though. When we started liquid cooling it,
like I mentioned earlier, right?

**[00:32:49]** It was an energy efficiency story
and in Munich, the customer also had an expectation of minimizing
water consumption, so they had

**[00:32:58]** a dry cooler in on the roof. The dry cooler has an expectation
of operation temperature.

**[00:33:05]** I mean, even though it's Munich in August,
it does gets warm, but, so we designed

**[00:33:11]** our platforms for W45 back in the day. Now we have a TCS at 45.

**[00:33:19]** I mean that W45 is difficult to maintain. So we have a W40 based cooling
solution and we are deployed it,

**[00:33:28]** essentially in a whole bunch of our
customer environments around the planet.

**[00:33:34]** Though what we are also seeing
is, customers who are retrofitting

**[00:33:40]** their air pool layer centers
that hatch chill water to liquid. Well, they're not deploying a dry cooler.

**[00:33:47]** They're using the existing data
central facility infrastructure. And in that environment
it's W 18 or 17, right?

**[00:33:56]** So that's the water temperature
coming into the, racks or the CDUs.

**[00:34:02]** So BCRH arrange though, with clarity,
the expectation of the higher water

**[00:34:10]** temperature stems from an expectation
of optimizing for energy resources,

**[00:34:15]** which includes water, right? I mean, water is a scarce normality. Yeah. So the more you can use free cooling
or a dry cooler, right, you can maybe

**[00:34:26]** get better energy for compute or token
creation, which is the metric today.

**[00:34:33]** Right. So it, so it's interesting to know
that you, you know, you guys have been doing 45 C water and 40 for,
for a while, so that's fantastic.

**[00:34:44]** And in a term, and this is in my learning,
I'm gonna ask you a question, Greg. You know, I hear about a CDU and
especially when you're trying to

**[00:34:52]** get as much free cooling as you can. You know, there's a talk about,
oh, the approach temperature across the heat exchanger.

**[00:34:57]** What's that mean in, in the system design? Yeah, so approach temperature delta, ATD
is what it's typically referred to, is

**[00:35:05]** kind of a, it's a temperature difference. So it's one temperature minus another
to give you some number, right?

**[00:35:10]** And that's gonna be your water. Or I guess in, in some systems,
obviously if it's air cold through

**[00:35:16]** your fluid, for instance, your fluid
going to the IT equipment, minus the

**[00:35:22]** incoming facility, water temperature. So if you have, you know, a cold
water temperature coming in here and

**[00:35:27]** a slightly warmer temperature going
to the equipment, that difference between those temperatures is, is
your approach temperature, delta, ATD.

**[00:35:35]** And so kind of getting into
the whole ecosystem, right? The chip to chiller story is we have
heat at the chip and it has to go

**[00:35:41]** through all these different stage
gates to get out to the air and go off into the atmosphere, right?

**[00:35:46]** And the chiller discussion and,
and dry cooler between those two. Right? So, each step, you're gonna
have another ATD, right?

**[00:35:54]** So between chip from that fluid and the
CDU going into the facility water, there's gonna be a bump in temperature from the
facility water to the refrigerant loop

**[00:36:02]** in, if it was a cooler, there's gonna
be another jump, jump in temperature. And so as you, in typical trends, right,
I don't want to, you know, name specific

**[00:36:11]** numbers, but if you can decrease that ATD,
you're gonna in increase your efficiency.

**[00:36:17]** Because, you know, you
have less of a bridge of a temperature gap to go to, right? Mm-hmm.

**[00:36:23]** And so the 45 degree discussion is
great because then you can run your

**[00:36:29]** system at a hotter temperature. And then as you step up to hotter and
hotter temperatures, then your final kind of, I don't know necessarily
ATD, but your difference between your

**[00:36:36]** ambient temperature outside and your
heat dissipation is gonna be greater and greater, so you have more capability to
dissipate that heat and the environment.

**[00:36:44]** Right? And that's the whole goal, goal of
trying to go to dry coolers, right? You don't need a refrigeration cycle. You could just dump the heat out for free.

**[00:36:50]** Right. So I guess, does that answer
the question or do we wanna go? Yeah, yeah.

**[00:36:55]** Feels like I ramble a little bit, but. Yeah. Let's, for a CDU very quickly, like what's
a typical, like the temperature difference

**[00:37:02]** or that approach temperature built up? Sure. Yeah. Within a CDE between the
facility water and the TCS.

**[00:37:08]** What, what are those? What's that number typically? Yeah. So, a good metric to go by, so
actuary very recently, so SSPC127

**[00:37:15]** is a special group, devising the
method of tests for CDUs and other components within the data center.

**[00:37:21]** So they've recently put out an
addendum that qualifies that method of tests to be set at four degrees ATD.

**[00:37:28]** So that's gonna be four degrees between
your incoming water to your ITE versus your cold water on the facility side.

**[00:37:34]** That's not to say that all CDUs have
to operate at four and they don't. Right? Yeah. So there, there can be trade offs either
way, which direction you want to go.

**[00:37:41]** There's this, but in general, as you want
to shrink your ATD, yeah, that's going

**[00:37:47]** to affect your CDU designs for sure. Yeah. Yeah. But four four would be typical, you
know, based on natural point of view.

**[00:37:54]** Okay. Okay. Alright, so we're getting towards
the end and I'm gonna do, some

**[00:37:59]** quick wrap up questions for you. So, if you know, I'm gonna have two
separate questions, you know, and if

**[00:38:09]** you're talking to, so if you were talking
to a designer, that might be designing their first liquid cooling system, you
know, what advice would you give them?

**[00:38:18]** And, Vinod, I'll start with you. Yeah. So what I would say is if the person has
never deployed these kind of solutions

**[00:38:28]** before, I would say startup by reading the
ASHRAE and OCP guidelines, start there.

**[00:38:36]** I mean that because that's
sort of vendor agnostic and you know, I'm all for that, right?

**[00:38:41]** So once you understand what the,
data, certainty, and the knowledge

**[00:38:48]** is in the industry, right? The terms, the expectations,
the way to manage it.

**[00:38:53]** Once you do that, understand what the
lifetime of your installation wants to be.

**[00:39:00]** Are you planning this for, you know,
one generation, three generations of products and sort of what is the level
of engagement that you want from the

**[00:39:10]** server vendor next to help engage with
you to understand what that solution

**[00:39:17]** path might be based on their technologies
in your data center environment. Right?

**[00:39:23]** At the moment, I think all the tier
one server vendors are aligned to

**[00:39:29]** deliver essentially reliable high
performance computing solutions. I believe that at Lenovo
restarted that journey.

**[00:39:37]** So in that model where you are following
industry standards and then engaging

**[00:39:43]** with the server vendor together, you
will basically find the right solution

**[00:39:48]** path identified for your data center
environment, whether it's the new one

**[00:39:54]** or an existing one, because we have
seen examples where, at Lenovo, we

**[00:39:59]** have engaged with a lot of enterprise
customers that they have redone the air pool environments with liquid cooling,
and that transition has gone smoothly.

**[00:40:09]** It has be done right, and, you know,
good conversations, but, it works.

**[00:40:15]** Yeah. That's great to hear. Yeah. Okay. Greg. Yeah, so I, I would say first off,
congrats on making the right steps towards

**[00:40:22]** improving your data center efficiency. Right. But I think, all jokes aside, I
think the best thing I can, you know,

**[00:40:28]** there, there's two things, right,
is one, just communication, right? As Vinod mentioned, you know, there's,
this is a, a wide system and there's a

**[00:40:35]** lot of different moving parts, and so
first of all, reading up on your ASHRAE and your OCP guidelines is a great start
to kind of get an understanding of, of

**[00:40:43]** what a liquid, liquid cooling system is. But then when you go to deploying there,
there's the server manufacturer, there's

**[00:40:49]** the CDU manufacturer, there is the
facility that's managing the data center. It's gonna be your, your chiller
maintenance, all these different, you

**[00:40:56]** know, aspects that are, might be in
different groups and might get siloed as far as, you know, who's talking to who.

**[00:41:01]** And so kind of encompassing the
entire system and thinking about all the different levels of that
and making sure that you're getting

**[00:41:07]** all those parties talking together. It is the best way to deploy, you
know, and make the right solution.

**[00:41:13]** Fantastic. I did mention two things. Yeah. So then I think the second thing,
just kind of a general term, just to kind of throw this out there.

**[00:41:19]** So LPM per kilowatt, that's kind of a,
an industry standard way to talk about if

**[00:41:24]** you're gonna deploy a, say, 500 kilowatt
system that's relatively small these

**[00:41:30]** days, but let's say that's, you know, a
couple racks, something like that, right? You can generally assume that 1.5
LPM per kilowatt gives you the total

**[00:41:38]** flow rate that you might need, right? And so that, that's kind of a general
guideline for sizing your, your total

**[00:41:43]** CDU system, that kind of for your power. That's not to say that your
power might be constant, right? You might be upgrading servers over
time, so maybe you want to consider

**[00:41:52]** what, again, as Vinod said, the total
lifecycle and how much power you might be, you know, consuming in 5
years, 10 years on the same system.

**[00:41:59]** Yeah, that's great. I love, I love thumb rules, 1.5, and
I'll specify liters per minute per kW.

**[00:42:06]** PM. Yes. Yeah. Thank you. It's okay.

**[00:42:11]** all right, the last question I have. All right, so we've gone through design. Now you're talking to a data center
operator who's installing and operating

**[00:42:20]** their first liquid cooled system. What advice would you have for them? So again, I'll start with Vinod.

**[00:42:26]** So assuming that the gear is on
site right, the first thing that you want to do is establish the
cleanliness of your TCS loop.

**[00:42:36]** Assuming you have it right. So this is absolutely critical because
if that's not done right, the first

**[00:42:43]** step is not done right, you will then
cascade and, create a whole set of

**[00:42:48]** problems that you may not know at time
B zero that will show up over time. So typically what that means is making
sure they either the, solution provider

**[00:43:00]** may not be the server vendor, right? Sometimes they, they're
secondary providers, right? Responsible for the maintenance have
documented to you that the servers and

**[00:43:10]** the CDUs are, will be connected to a loop. There are going to be of perfect
chemistry, but once you establish that

**[00:43:17]** and you have the servers that have
been sort of well vetted or calibrated,

**[00:43:24]** you know, from the supplier connected,
you want to be sure that you have the

**[00:43:29]** right fill and drain process because
in some cases, you know, these servers will not come prefilled with EM PG 25.

**[00:43:36]** They may be deploy on site. Water clearly is deployed on site, right?

**[00:43:41]** So you wanna make sure you're doing
the right fill and drain process. And then when you power the servers
on, once you establish the fluid

**[00:43:51]** flow rates and temperatures are
set, right, you want to then establish using your baseline tests.

**[00:43:57]** Things like heat capture. Do you know the amount of heat the servers
are racks are participating from your CDU?

**[00:44:06]** And the m-dot, cp, delta T that Greg
mentioned, do you have certainty

**[00:44:12]** that the amount of heat that, you
knew that the liquid is gonna be captured is actually being seen

**[00:44:19]** running workloads, right, at time to load. Because once you establish that, you
know the flow rates are right, you

**[00:44:25]** know that the dps are right and you
know, as we start, talk about the fluid chemistry is right and once that's
done right, then it's a simple fluid

**[00:44:35]** maintenance, you know, cycle that you
follow and you have a clean, reliable, high performance computing environment.

**[00:44:42]** Fantastic. Great, and Greg? Yeah. So I guess my response to the previous
question is probably, you know, analogous

**[00:44:49]** here too about communication, right? So to Vinod's point about, you know,
making sure we have a very clean fluid,

**[00:44:56]** I'll use a, an anecdote horror story
that I've heard of where there wasn't communication and we, there was TCS
loops that were installed, right?

**[00:45:03]** That might've been open to the air. And then there was, maybe a facility
contractor that was doing some

**[00:45:09]** drywall work in the same room. Obviously not a very clean environment. But that could have been avoided
by, again, communication between all

**[00:45:15]** parties in the data center design. So just making sure that we're keeping
that communication line open so that

**[00:45:21]** all facilities know when stuff is
being done to keep that fluid clean.

**[00:45:26]** As far as the second point, kind of
getting back to the circuit analogy, the V=IR, if you look at that analogy
that can be used for, you know,

**[00:45:36]** parallel and series resistances. And if you consider the, the flow
resistance of the network as well,

**[00:45:42]** that the, another, that's actually
a third analogy that fits the same V=IR there, there's thermal, there's
current and then there's, you know,

**[00:45:49]** flow network modeling, which would be
your pressure drop in designing how much flow goes to different servers, right?

**[00:45:55]** And so if you have a, you know, raw
base CDU that's providing, fluid,

**[00:46:00]** then you have two different servers
of two different, flow resistances. They're gonna get two
different amounts of flow.

**[00:46:06]** Right. Because you're, the flow is gonna take
the path of least resistance, right? So this, key point of advice is to
consider that when you're designing

**[00:46:13]** your, your system and typically
you're want, wanting to use equal servers of equal pressure drop
so that they all get equal flow.

**[00:46:21]** Right? As a tip. Yeah. Fantastic. Should anybody be be afraid
of deploying liquid cooling?

**[00:46:29]** No. Perfect. I think we're well past that. Yeah. That's great.

**[00:46:35]** So, I want to thank you very,
very much for your time. I, I know you're both very busy
and, and again, you, having people

**[00:46:44]** that have been there and done that
for quite a number of years is, is super valuable to the industry.

**[00:46:51]** So, you know, again, thank
you for your time today. Absolutely. Alright, you're welcome. Happy to help advocate
and keep things simple.

**[00:46:58]** Great. All right. Thank you.