Why are AI data centers using so much electricity? - Sajan Saini · TED-Ed
了解AI数据中心的能耗问题，学习与科技、环境相关的英语表达，适合练习科技类话题和复杂句型。
Since 2019, over a thousand imposing warehouse campuses have popped up around the world.
自2019年以来，全球各地涌现出了超过一千座规模庞大、气势恢宏的仓储设施。
Each complex covers nearly 100,000 square meters, uses as much water as a small town, and as much electricity as a small city.
每个建筑群占地面积接近10万平方米，耗水量相当于一个小城镇的用水量，耗电量则相当于一个小城市的用电量。
These are AI data centers— facilities for running and training Large Language Models like Claude and ChatGPT.
这些是人工智能数据中心，用于运行和训练诸如 Claude 和 ChatGPT 之类的大型语言模型。
We’ve had similar data centers dedicated to operating the internet since the 2000s,
自21世纪初以来，我们就一直拥有专门用于运营互联网的类似数据中心。
but these customized complexes come with a whole new suite of challenges and possibilities.
不过，这些经过定制的综合性解决方案（或系统）带来了一整套全新的挑战与可能性。
AI data centers are generally organized into several zones.
人工智能数据中心通常被划分为多个区域（或“分区”）。
The Network Operations Center manages data traffic and power consumption.
网络运营中心负责管理数据流量和电力消耗情况。
Power systems including transformers, gas generators, and battery-powered backups ensure uninterrupted electricity flow.
电力系统包括变压器、燃气发电机以及由电池提供能量的备用电源，这些设备能够确保电力供应的连续性（即不会中断）。
And a connectivity hub called the Meet-Me-Room directs incoming data into the heart of the data center: the server room.
此外，还有一个名为“Meet-Me-Room”的连接枢纽，它将传入的数据导向数据中心的核心区域——即服务器机房。
When you ask an LLM a question, your prompt enters one of these server rooms.
当你向大型语言模型（LLM）提问时，你输入的问题会被发送到其中一个服务器机房中进行处理。
It's too complicated to explore how that model responds right now, but it involves doing tons of math.
目前要研究该模型是如何响应的实在太复杂了，但这需要大量的数学计算。
To manage these calculations, a data center relies on thousands of server racks, each stacked with trays containing server boards that house memory chips and GPUs.
为了处理这些计算任务，数据中心需要依赖数千个服务器机架；每个机架上都堆放着多个托盘，而这些托盘内装有装有内存芯片和显卡的服务器主板。
These powerhouse microchips use thousands of tiny computing cores working in parallel to do trillions of mathematical operations per second.
这些高性能微芯片配备了数千个并行工作的微型计算核心，每秒能够完成数万亿次数学运算。
While most prompts only require a few GPUs to answer, data centers already field tens of thousands of prompts a second.
虽然大多数问题只需要几块 GPU 就可以回答，但数据中心每秒钟仍要处理成千上万条这样的请求。
And that rate is growing fast.
而且这个增长率正在迅速上升。
Plus, training LLMS requires hundreds of thousands of GPUs working together, making server grids like these essential to get an AI up and running.
此外，训练大型语言模型（LLMs）需要数十万个GPU协同工作，因此这样的服务器集群对于AI系统的顺利运行至关重要。
However, all this energy-intensive work leads to a major problem: heat.
然而，所有这些耗能巨大的工作都会引发一个严重的问题：热量。
Even in climate-controlled server rooms, GPUs get hot enough to melt server boards.
即使在温度受控制的服务器机房中，显卡也会发热到足以熔化服务器主板的程度。
And the copper wires connecting these chips constantly dissipate energy, using more electricity and generating additional heat.
连接这些芯片的铜线会不断消耗能量，从而消耗更多的电能并产生额外的热量。
Fortunately, engineers are on the job.
幸运的是，工程师们已经在着手解决这个问题了。
To cool chips, they’re using tried-and-true solutions like water towers that absorb and evaporate heat alongside new approaches like piping chilled water behind server boards,
为了给芯片降温，他们采用了经过验证的有效方法（例如用于吸收和蒸发热量的水塔），同时也尝试了一些新的方法（比如将冷却后的水通过管道输送到服务器主板后方）。
or dunking entire trays into non-conductive cooling liquids.
或者将整个托盘浸入非导电的冷却液中。
At the same time, they're replacing inefficient copper wires that transmit data through electricity with optical fibers that transmit data through light.
与此同时，他们正在用光纤取代那些通过电来传输数据、效率低下的铜线；光纤则是通过光来传输数据的。
At the top of many racks, optical transceivers convert electrical data to light signals, which travel through optical fibers with higher data fidelity than copper
在许多机架的顶部，光收发器会将电信号转换为光信号；这些光信号通过光纤传输，其数据传输的保真度远高于铜缆。
and near-zero power loss.
功率损耗几乎为零。
And as photonics technology gets further integrated inside servers, it can solve another major problem.
随着光子技术在服务器内部的集成程度不断提高，它还能解决另一个重大问题。
GPUs only have so much space for copper connections, limiting how much data they transmit at once.
GPU内部的铜质连接线路所占空间有限，这限制了它们同时能够传输的数据量。
But integrated photonic chips can use different wavelengths of light to send multiple streams of data over a few optical lines.
不过，集成光子芯片可以利用不同波长的光，通过几条光缆同时传输多路数据。
This same tech could even let GPUs interact with memory chips on neighboring trays and racks—
同样的技术甚至可以让图形处理器（GPU）与相邻托盘或机架上的内存芯片进行交互。
spreading out their heat load and allowing bigger LLMs to use less power.
通过分散热负荷，使得大型语言模型（LLMs）能够使用更少的电力来运行。
While some experts believe integrated photonics will reduce data center power use by nearly 50%, this still might not be enough.
虽然一些专家认为集成光子技术能够将数据中心的能耗降低近50%，但这可能仍然不够。
GPUs run at roughly 16 light bulbs worth of energy.
GPU的运行所消耗的能量大约相当于16个普通灯泡的能耗。
Multiply that by every GPU in a data center, and you're looking at enough energy to power 300,000 homes.
将这个数值乘以数据中心中所有 GPU 的总功耗，结果得到的能量足以供应 30 万户家庭的用电需求。
Companies are also pursuing ever-smarter AI, which, so far, has meant ever-larger AI.
各公司也在不断研发更智能的人工智能技术，而到目前为止，这种技术的发展趋势就是人工智能系统的规模越来越大（即计算能力、处理能力等持续增强）。
Even with increasing efficiencies, larger models require bigger data centers and more electricity.
尽管效率不断提高，但更大规模的模型仍然需要更庞大的数据中心和更多的电力支持。
Some estimates predict that by 2028, US data centers alone could be using as much energy as eight New York Citys.
一些估计数据显示，到2028年，仅美国的数据中心所消耗的能源就相当于8个纽约市的能源总量。
In fact, some AI companies have already proposed pairing next-gen data centers with their own personal nuclear power plant.
事实上，一些人工智能公司已经提出将下一代数据中心与他们自己拥有的核电站相结合的想法。
Clearly, we don’t have enough energy to support this rapid expansion, so we either need to hit the brakes or find ways to compute smarter.
显然，我们没有足够的能源来支持这种快速的扩张；因此，我们要么必须“踩刹车”（即停止这种扩张），要么就必须找到更高效、更智能的计算方法。
For example, while current LLMs are designed to answer every possible question, you don’t need a model that knows Shakespeare to help you fix your dishwasher.
例如，虽然当前的 large language models（LLMs）被设计用来回答所有可能的问题，但你并不需要一个精通莎士比亚作品的模型来帮你修理洗碗机。
We could design topic-specific “small language models” stored in micro-data centers.
我们可以设计针对特定主题的“小型语言模型”，并将这些模型存储在微型数据中心中。