dc-thermal
Menu菜单
← All writing← 全部文章

Where Does Air Cooling End? An HVAC Engineer's Reading of the AI Rack风冷的终点在哪里？一个暖通工程师看 AI 机柜

Willow LeiIndustry insight行业洞察

Ten years ago, 12 kW per rack was already considered extreme density. I worked on data center projects that cooled racks like that with two-phase CO₂: the servers were still cooled by their own fans, and the CO₂ carried the heat away in rack-level heat exchangers, so from the server's point of view it was still air cooling. Back then it felt like the edge. Today the question I hear most from colleagues is simple: at what rack power can we no longer use air?

I have looked at the public numbers, and my answer is that, for the rack, air cooling has no fixed end point. It moves. At some point air stops being the main heat path, but air systems stay in the data hall long after that. The real changes happen in two places: inside the rack, and in the central plant outside the hall.

AI racks keep getting denser

First, a question of basis. Nearly all the electrical power drawn by IT equipment becomes heat, so HVAC design uses rack power as the heat load. But rack power is quoted in at least three ways: a server's maximum AC input, a rack's design power, and a cabinet's thermal design power (TDP). All three can estimate the heat load, but they are not equally conservative, and mixing them is the most common mistake I see. For example, NVIDIA's "designed rack power" for GB300 NVL72 is 135 kW, the budget when the infrastructure "is provisioned for the full theoretical load"; Lenovo calls the same 135 kW the TDP and notes peaks of up to 155 kW. Neither design power nor TDP is an absolute maximum.

With that in mind, here are the public numbers from NVIDIA's own documents (basis in brackets):

  • DGX H200: 10.2 kW maximum AC input per system; the reference architecture places two per rack, so about 20 kW per rack (server maximum × units, our arithmetic, excluding network and storage).
  • DGX B200: 14.3 kW maximum per system, two per rack, about 29 kW (same basis).
  • DGX B300: 15 kW maximum per system, and up to four per rack "if there is additional power and cooling capacity", up to about 60 kW (same basis).
  • GB200 NVL72: 120 kW designed rack power; GB300 NVL72: 135 kW (designed rack power).
  • NVIDIA's DSX reference design: cabinet TDP from 198 kW up to 330 kW for Vera Rubin NVL72 (cabinet TDP).

The cooling specifications of these 8-GPU DGX systems are air specifications, given as airflow and temperature — for example, 1,550 CFM and an operating temperature of 10–35 °C for B200. From GB200 NVL72 on, NVIDIA's rack-scale systems themselves are designed for liquid cooling. In other words, within the same Blackwell generation, 8-GPU servers are pushing air to about 60 kW per rack while the rack-scale systems have gone straight to liquid.

An estimate of the air-cooled rack limit

My judgement is that for today's air-cooled AI servers, air remains a comfortable main heat path at roughly 20 to 60 kW per rack. NVIDIA itself attaches a condition to the four-per-rack B300 layout: additional power and cooling capacity. That condition is exactly where room-side engineering gets hard — airflow per rack, containment, rear-door heat exchangers and fan power.

I do not treat this as a fixed threshold. The real limit depends on the server model, the airflow and temperature rise, the climate and the rating curves of the actual equipment. A number without those conditions is a slogan, not an engineering limit.

Air does not disappear; its share shrinks

Liquid-cooled racks still put some heat into the air. NVIDIA calls GB300 NVL72 a "fully liquid-cooled, rack-scale architecture", yet Lenovo describes the same system as about 90% liquid-cooled and 10% air-cooled. "Fully liquid-cooled" describes how the compute is cooled; it does not mean zero heat to air. Ten percent of 135 kW is roughly 13.5 kW per rack — about two-thirds of the maximum input of a whole air-cooled H200 rack. The hall still needs fan walls or CRAH units, just for a smaller share.

That share can go to zero. NVIDIA now says its MGX racks can be 100% liquid-cooled. When that becomes the normal deployment, the hall air system becomes a comfort and residual-load system rather than the IT cooling system.

The real change is in the plant

The bigger shift is temperature. NVIDIA's reference design uses a 45 °C liquid-cooling design point, which "expands the operating window for rejecting facility heat without full mechanical chilling". In NVIDIA's example, 41 °C facility water goes to the CDUs, which supply 45 °C coolant to the racks.

For an HVAC engineer, that changes the plant: more dry coolers, fewer chiller running hours, and a dual-temperature central plant that serves the liquid loop and the remaining air load separately. ASHRAE's water classes, from W17 to W45 and W+, give us the vocabulary for this.

Warm water is not free, though. ASHRAE's own case study shows that moving from W27 to W45 water can cut cold-plate cooling capability for that server to roughly what a well-designed air system could do. Every degree of water temperature is a trade between chiller savings and chip-side margin.

Where I could be wrong

Two things could prove this view wrong. First, air-side engineering keeps improving: in retrofits, rear-door heat exchangers and air-assisted liquid cooling may keep air as the main path for longer than I expect. Second, the next racks are not yet clearly specified. The widely repeated "600 kW Kyber rack" is not what NVIDIA's current documents say; they describe Kyber as a 144-GPU rack and give no rack power. Until vendors publish rack power with its basis, any forecast — including mine — is provisional.

I will review this view at the end of 2027, when the next rack generation should be shipping.

Personal view. Public sources only; numbers keep their stated basis. No employer view is implied.

Sources

Sources checked 2026-10-11.

十年前，每个机柜 12 kW 已经算极高密度。我当时参与过用两相 CO₂ 冷却这类机柜的数据中心项目：服务器本身仍然靠风扇吹风散热，CO₂ 在机柜级的换热器里把热量带走，所以从服务器的角度看，它仍然是风冷。那时觉得已经到了边缘。现在同行问我最多的问题很简单：机柜功率到多少，就不能用风冷了？

我看了公开数据，我的回答是：对机柜来说，风冷没有一个固定的终点数字，它是在“移动”的。到了某个点，空气不再是主要的散热路径；但在那之后很长时间，空气系统仍然留在机房里。而真正发生大变化的地方有两个：一个在机柜里面，另一个在数据中心外围的冷站。

AIDC 数据中心的机柜热密度越来越高

先说一个口径问题。进入 IT 设备的电功率几乎全部变成热，所以暖通设计用机柜功率来估算散热负荷。但机柜功率至少有三种不同的说法：服务器的最大交流输入、机柜的设计功率、整柜的热设计功率（TDP）。它们都能用来估算热负荷，但保守程度不同，混在一起比较是我看到最常见的错误。例如 NVIDIA 给 GB300 NVL72 的“设计机柜功率”是 135 kW，指的是按“全部理论负荷”配置供电时的预算；Lenovo 把同一个 135 kW 称为 TDP，同时注明峰值可达 155 kW。所以设计功率和 TDP 都不是绝对最大值。

在这个前提下，NVIDIA 自己的文档给出的公开数据如下（括号里是口径）：

  • DGX H200 单台最大交流输入 10.2 kW，参考架构每柜 2 台：每柜约 20 kW（服务器最大输入 × 台数，是我们的算术，不含网络和存储）。
  • DGX B200：单台最大 14.3 kW，每柜 2 台，约 29 kW（同上）。
  • DGX B300：单台最大 15 kW，“在有额外供电和冷却能力时”每柜最多 4 台，最高约 60 kW（同上）。
  • GB200 NVL72 设计机柜功率 120 kW；GB300 NVL72 为 135 kW（设计机柜功率）。
  • NVIDIA DSX 参考设计中，整柜 TDP 从 198 kW 到 Vera Rubin NVL72 的 330 kW（整柜 TDP）。

前三种 8 卡 DGX 服务器的散热规格都是按风冷给出的：用风量和进风温度描述，例如 B200 风量 1,550 CFM、运行温度 10–35 °C。而从 GB200 NVL72 开始，NVIDIA 的整机柜系统本身就是按液冷设计的。也就是说，在同一代 Blackwell 里，8 卡服务器还在把风冷推到每柜约 60 kW，整机柜系统已经直接走向液冷。

风冷的单机柜极限预估

我的判断是：对今天的风冷 AI 服务器，风冷作为“舒服的”主散热路径，大约在每柜 20 到 60 kW 之间。NVIDIA 自己给每柜 4 台 B300 的布置加了条件：需要额外的供电和冷却能力。这个条件就是房间侧工程变难的地方：每柜风量、封闭通道、背板换热器和风机功耗。

我不把这当成固定门槛。真正的上限取决于服务器型号、风量和温升、气候，以及实际设备的额定曲线。不带这些条件的数字只是口号，不是工程极限。

但风冷不会消失，只是比例小了

液冷机柜仍然有一部分热量进入空气。NVIDIA 把 GB300 NVL72 称为“全液冷”的整机柜架构，但 Lenovo 对同一系统的描述是约 90% 液冷、10% 风冷。“全液冷”说的是计算部分的散热方式，不等于没有热量进入空气。135 kW 的 10% 约为每柜 13.5 kW，相当于一整柜风冷 H200 最大输入的三分之二左右。机房仍然需要风墙或精密空调，只是承担的份额变小了。

这个份额可以降到零。NVIDIA 现在表示其 MGX 机柜可以 100% 液冷。当这成为常规部署时，机房空气系统就从 IT 冷却系统变成了处理舒适和残余负荷的系统。

真正的变化在冷站

更大的变化是温度。NVIDIA 的参考设计采用 45 °C 液冷设计点，它“扩大了不需要完全机械制冷就能排出设施热量的运行范围”。在 NVIDIA 的示例中，41 °C 的设施水送到 CDU，CDU 再向机柜供应 45 °C 的冷却液。

对暖通工程师来说，这句话改变了冷站：更多干冷器、更少的冷机运行小时，以及分别服务液冷回路和剩余空气负荷的双温冷站。ASHRAE 的水温等级 W17 到 W45 和 W+，就是描述这件事的语言。

但温水不是免费的。ASHRAE 自己的案例显示，从 W27 提高到 W45，对那种服务器的冷板冷却能力可能降到与一个设计良好的风冷系统相当的水平。水温每提高一度，都是在冷机节能和芯片侧裕量之间做取舍。

我可能错在哪里

有两件事可能证明我错了。第一，风侧工程还在进步：在改造项目里，背板换热器和风液混合方案可能让空气作为主路径的时间比我预期的更长。第二，下一代机柜的规格还不清楚。广为流传的“600 kW Kyber 机柜”并不是 NVIDIA 当前文档的说法；NVIDIA 把 Kyber 描述为 144 GPU 的机柜，没有给出机柜功率。在厂商公布带口径的机柜功率之前，任何预测（包括我的）都是暂时的。

我会在 2027 年底下一代机柜开始出货时回看这个判断。

个人观点，只使用公开资料，数字保留原始口径，不代表雇主立场。

资料来源

资料核查日期：2026-10-11。

All writing →全部文章 →