IT 领导者必须评估当前的数据中心，以跟上当今不断变化的运营需求。
在人工智能 （AI） 时代，企业如何评估其现有的数据中心设计是否能够充分利用运行 AI 所需的现代要求？在 IT 领导者制定其 AI 战略并评估其基础架构的前景时，需要考虑一些主要因素。
这篇博客探讨了： 数字化转型网（www.szhzxw.cn）
- 什么是传统 IT 基础架构？
- 如何将新的 AI 设备与现有基础设施集成。
- 评估数据中心设计和传统基础设施。
- 数据中心改造的艺术。
尽管 AI 工作流程围绕独特的电源和冷却需求提出了新的挑战和问题，但 IT 领导者应评估其数据中心设计状态，以适应新兴和不断发展的现代需求。
一、什么是传统 IT 基础架构？
识别传统基础设施既是直觉，也是经验。从IT设备的角度来看，我们可以假设任何不处于最前沿的东西都是遗留的。然而，这往往不是真的。 数字化转型网（www.szhzxw.cn）
世界上许多 IT 系统都无法在最新、最出色的硬件上运行。由于从超大规模云到小型企业的每个人的典型预算、支出和设备更新周期，这种情况将继续下去。
即使在人工智能时代，也不是每个机架都能消耗 100 kW 的功率或需要液体冷却。装满网络、内存聚合或存储设备的机架可能仍低于 15 kW，并且依赖于空气冷却。
仅根据其功耗将 IT 基础设施归类为遗留或非传统基础设施变得具有挑战性。各种行业基准测试表明，新一代中央处理器 （CPU）、图形处理单元 （GPU）、网络设备和其他 IT 基础设施资产明显快于其前代产品，但仅凭这一点通常不足以将现有设备指定为传统基础设施。
最好的测试是确定当前的基础设施是否以新一代设备不会阻碍组织的发展和运营活动。
如果是这样，则应将其归类为遗留基础设施。
二、将 AI 与现有 IT 基础设施集成
就 IT 设备而言，我们可以将 AI 集成为利用现有服务器及其支持设备来执行新的 AI 功能，或者使用新的 AI 特定设备增强部署的硬件以执行新的 AI 功能。 数字化转型网（www.szhzxw.cn）
后者的一个例子是利用现有的基于 CPU 的服务器机架，并添加两个新的基于 GPU 的服务器，以提供更多的并行计算能力，以便向公司的内部用户启动聊天机器人。
这似乎比适应新的 AI 高密度部署更容易，但它带来了三组挑战：
- 将基于 GPU 的服务器添加到原本机架密度较低的通道中可能会产生建筑物的冷却系统最初设计无法处理的热点。
- 它可能会在整个设施中造成不均匀的电力负载，并导致需要重新分配备用电源资源。
- 这可能会导致网络拥塞，因为新设备会成倍增加每个机架传输的数据。
这些因素可能会给数据中心带来新的压力，您应该将其视为 IT 堆栈本身的一部分。
三、评估数据中心设计和传统基础设施
数据中心与部署在其中的服务器一样，都是 IT 基础设施的一部分，因此我们应该考虑遗留基础设施的概念如何也适用于数据中心设施。 数字化转型网（www.szhzxw.cn）
从技术角度来看，数据中心行业不是春鸡。仅Digital Realty就支持全球约2.4吉瓦的客户IT设备，这并非一蹴而就。
自 2004 年成立以来，我们每年都在逐步增加我们的全球数据中心容量，并且这些设施中的所有客户设备都不会消失。许多组织每三到五年更换一次所有服务器，但有些服务器可能部署长达八年。更换设备时，将分阶段完成，以便组织的应用程序在不发生实际停机的情况下运行。
这意味着数据中心始终处于开启状态。数据中心运营商不能简单地将所有客户的 IT 设备取出，对设施进行大规模升级，然后将其全部放回原处。随着时间的流逝，数据中心中的客户设备组合通常包含一些遗留设备和一些非遗留设备。 数字化转型网（www.szhzxw.cn）
此外，随着数据中心设施本身的老化，其自身的一些特性（例如气流设计、地板结构和对液体冷却的支持）可能并不适合客户想要部署的所有设备。
例如，许多数据中心设施都使用活动地板设计——AI 设备不仅在功耗方面推动了更高的机架密度，还因为它们的重量。在某些情况下，这些机架可能需要坚固的混凝土板地板。
这意味着对于某些用例，某些数据中心可以符合我们对传统基础设施的定义。
然而，一个精心设计的数据中心在随时间推移的升级方面比跨多个机架的服务器或一组 IT 设备要灵活得多。数据中心的使用寿命可能超过 15 到 20 年，具体取决于运营商随着时间的推移对其进行设计、改造和模块化的表现。 数字化转型网（www.szhzxw.cn）
人工智能促使机架密度和其他影响数据中心的要求发生了翻天覆地的变化。通常，数据中心运营商可以升级部分设施以适应这些新需求。
四、数据中心设计改造的艺术
这个过程被称为改造，改造的艺术是数据中心运营商在为当前和未来几代服务器和其他IT基础设施设计数据中心时如何有效的关键组成部分。想象一下，一个较旧的数据中心设施最初设计为平均每个机架 10 kW。随着人工智能的出现，该设施有望支持每个机架 100 kW 的功率，而无需完全关闭和从头开始重新设计。 数字化转型网（www.szhzxw.cn）
随着时间推移，支持数据中心的这些类型的更改的灵活性是我们设计和运营数据中心的关键部分。例如：
- 如果不再需要活动地板，则可以将其填充。
- 在需要液体冷却的情况下，我们可以将管道从新的冷水机组和储液罐连接到机架。
- 在需要新的网络功能的情况下，我们可以引入额外的连接，并优化设施内部的所有网络资产以匹配。
Digital Realty首席技术官Chris Sharp，“Digital Realty模块化设计方法的优势”
如今，数据中心与 IT 堆栈的任何其他部分一样灵活、模块化且高度调整，可随着客户的需求而发展。在数据中心支持人工智能的要求无疑是具有挑战性的，我们分析了全球所有数据中心，以掌握如何发展我们的设计和运营，以适应不断发展的人工智能需求。 数字化转型网（www.szhzxw.cn）
通过面向未来的 AI 就绪型基础架构，为您的 IT 战略赋能。
英文原文：
Data center design in the age of AI: Integrating AI with legacy infrastructure
IT leaders must evaluate current data centers to keep pace with today’s ever-evolving operational needs.
In the age of artificial intelligence (AI), how can enterprises evaluate whether their existing data center design can fully employ the modern requirements needed to run AI? There are major considerations as IT leaders develop their AI strategies and evaluate the landscape of their infrastructure.
Though AI workflows raise new challenges and questions around unique power and cooling needs, IT leaders should evaluate the state of their data center design to fit the needs of emerging and evolving modern needs.
1. What is legacy IT infrastructure?
Identifying legacy infrastructure is part intuition and part experience. From the perspective of IT equipment, we could assume that anything that is not at the bleeding edge is legacy. However, this is often not true. 数字化转型网（www.szhzxw.cn）
Many of the world’s IT systems do not run on the latest and greatest hardware. This will continue due to the typical budget, spending, and equipment update cycles of everyone from the hyperscale cloud to small enterprise.
Even in the age of AI, not every rack will be drawing 100 kW or need liquid cooling. Racks full of network, memory aggregation, or storage appliances may still be below 15 kW each and reliant on air cooling.
It becomes challenging to classify IT infrastructure as legacy or not based on its power draw alone. Various industry benchmarks show that new generations of central processing units (CPUs), graphics processing units (GPUs), network equipment, and other IT infrastructure assets are significantly faster than their predecessors, but this alone is often not enough to designate existing equipment as legacy infrastructure. 数字化转型网（www.szhzxw.cn）
The best test is to decide whether current infrastructure is holding back the development and operational activities of the organization in ways that new generations of equipment would not.
If it does, it should be classified as legacy infrastructure.
2. Integrating AI with existing IT infrastructure
In the case of IT equipment, we can think of integrating AI as either utilizing existing servers and their supporting equipment to perform new AI functions or augmenting hardware deployed with new AI-specific equipment to perform new AI functions.
An example of the latter is taking an existing rack of CPU-based servers and adding two new GPU-based servers to provide more parallel computing muscle to launch a chatbot to a company’s internal users.
This may seem easier than accommodating a new AI high-density deployment, but it comes with three sets of challenges:
- Adding GPU-based servers to an otherwise low rack density aisle may create hot spots that the building’s cooling system was not originally designed to handle.
- It may create uneven power loads across the facility and lead to the need to reallocate backup power resources.
- It may lead to network congestion as the new equipment multiplies the data transferred per rack.
These factors can lead to new pressure on the data center, which you should consider as part of your IT stack itself. 数字化转型网（www.szhzxw.cn）
3. Evaluating data center design and legacy infrastructure
The data center is as much a piece of your IT infrastructure as the servers that you deploy in it, and so we should consider how this concept of legacy infrastructure applies to the data center facility as well.
In technology terms, the data center industry is no spring chicken. Digital Realty alone supports around 2.4 gigawatts of customer IT equipment globally, and this didn’t happen overnight.
Since our founding in 2004, we have incrementally added to our global data center capacity each year, and all of the customer equipment in those facilities doesn’t go away. Many organizations replace all their servers every three to five years, but some servers may be deployed for up to eight years. When equipment is replaced, it’s done in phases, so that the organization’s applications operate without real downtime. 数字化转型网（www.szhzxw.cn）
What this means is that the data center is always on. A data center operator can’t simply lift all its customers’ IT equipment out, perform a wholesale upgrade of the facility, and then put it all back in. As time passes, the mix of customer equipment in the data center will typically contain some legacy and some non-legacy equipment.
Additionally, as the data center facility itself ages, some of its own characteristics, such as its airflow design, floor construction, and support for liquid cooling, may not be ideally suited to all the equipment that customers want to deploy.
For example, many data center facilities use a raised floor design— AI equipment drives higher rack densities not just in terms of power draw, but also because of their weight. In some cases, these racks can need a solid concrete slab floor. 数字化转型网（www.szhzxw.cn）
This means that for certain use cases, some data centers can fit our definition of legacy infrastructure.
However, a well-designed data center is far more flexible regarding upgrades over time than a server or set of IT equipment across multiple racks. A data center may last over 15 to 20 years, depending on how well the operator designs, retrofits, and modularizes it over time.
AI has prompted a sea change in rack densities and other requirements that impact the data center. Often the data center operator can upgrade parts of the facility to accommodate these new needs.
4. The art of the data center design retrofit
This process is known as retrofitting, and the art of the retrofit is a key component in how effective the data center operator is in designing data centers for current and future generations of servers and other IT infrastructure. Imagine an older data center facility originally designed for 10 kW per rack on average. With the emergence of AI, that same facility may be expected to support 100 kW per rack without the luxury of a total shutdown and ground-up redesign.
The flexibility to support these types of changes to the data center over time is a key part of how we design and operate our data centers. For example: 数字化转型网（www.szhzxw.cn）
- Where a raised floor is no longer needed, it can be filled in.
- Where liquid cooling is required, we can run piping from a new chiller unit and reservoir to the rack.
- Where new network capabilities are required, we can bring in additional connectivity and optimize all of the network assets inside the facility itself to match.
Digital Realty CTO Chris Sharp, “The Benefits of Digital Realty’s Modular Design Approach”
Today, the data center is as flexible, modular, and highly tuned to evolve with the needs of its customers as any other part of your IT stack is. The requirements to support AI in the data center are certainly challenging, and we analyze all our data centers globally to stay on top of how to evolve our design and operations to accommodate developing AI requirements.
Empower your IT strategy with future-proofed AI-ready infrastructure.
本文由数字化转型网（www.szhzxw.cn）转载而成，来源于CIO.com；编辑/翻译：数字化转型网宁檬树。
免责声明: 本网站(http://www.szhzxw.cn/)内容主要来自原创、合作媒体供稿和第三方投稿，凡在本网站出现的信息，均仅供参考。本网站将尽力确保所提供信息的准确性及可靠性，但不保证有关资料的准确性及可靠性，读者在使用前请进一步核实，并对任何自主决定的行为负责。本网站对有关资料所引致的错误、不确或遗漏，概不负任何法律责任。
本网站刊载的所有内容(包括但不仅限文字、图片、LOGO、音频、视频、软件、程序等) 版权归原作者所有。任何单位或个人认为本网站中的内容可能涉嫌侵犯其知识产权或存在不实内容时，请及时通知本站，予以删除。
