云端原来是重工业！为什么 AI 数据中心为何变成“耗电怪兽”
字幕摘录
| 时间 | 英文 | 中文 |
|---|---|---|
| 0:07 | Since 2019, over a thousand imposing warehouse campuses have popped up around the world. | 自2019年起,超过一千个强加的仓库校园在世界各地涌现. |
| 0:15 | Each complex covers nearly 100,000 square meters, uses as much water as a small town, | 每个建筑群占地近10万平方米 用水量和小镇一样多 |
| 0:22 | and as much electricity as a small city. | 电量和小城市一样大 |
| 0:26 | These are AI data centers, facilities for running and training large-language models | 这些是人工智能数据中心、运行和培训大语言模型的设施。 |
| 0:33 | 就像克劳德和ChatGBT | |
| 0:36 | We've had similar data centers dedicated to operating the internet since the 2000s, | 我们从2000年代起就有类似的数据中心 专门运行互联网, |
| 0:42 | but these customized complexes come with a whole new suite of challenges and possibilities. | 但这些定制的建筑群带来了一整套新的挑战和可能性。 |
| 0:50 | AI data centers are generally organized into several zones. | AI数据中心一般分为几个区. |
| 0:54 | The Network Operations Center manages data traffic and power consumption. | 网络操作中心管理数据流量和电力消耗. |
| 1:00 | Power systems, including transformers, gas generators, and battery-powered backups, | 动力系统,包括变压器、燃气发电机和电池动力备份, |
展开字幕全文（73 条）
| 序号 | 英文 | 中文 |
|---|---|---|
| 1 | Since 2019, over a thousand imposing warehouse campuses have popped up around the world. | 自2019年起,超过一千个强加的仓库校园在世界各地涌现. |
| 2 | Each complex covers nearly 100,000 square meters, uses as much water as a small town, | 每个建筑群占地近10万平方米 用水量和小镇一样多 |
| 3 | and as much electricity as a small city. | 电量和小城市一样大 |
| 4 | These are AI data centers, facilities for running and training large-language models | 这些是人工智能数据中心、运行和培训大语言模型的设施。 |
| 5 | 就像克劳德和ChatGBT | |
| 6 | We've had similar data centers dedicated to operating the internet since the 2000s, | 我们从2000年代起就有类似的数据中心 专门运行互联网, |
| 7 | but these customized complexes come with a whole new suite of challenges and possibilities. | 但这些定制的建筑群带来了一整套新的挑战和可能性。 |
| 8 | AI data centers are generally organized into several zones. | AI数据中心一般分为几个区. |
| 9 | The Network Operations Center manages data traffic and power consumption. | 网络操作中心管理数据流量和电力消耗. |
| 10 | Power systems, including transformers, gas generators, and battery-powered backups, | 动力系统,包括变压器、燃气发电机和电池动力备份, |
| 11 | ensure uninterrupted electricity flow. | 确保不间断的电力流通。 |
| 12 | And a connectivity hub called the MeetMe Room directs incoming data | 一个叫做MeetMe Room的连通中心 负责接收数据 |
| 13 | into the heart of the data center, the Server Room. | 进入数据中心的核心,服务器室。 |
| 14 | When you ask an LLM a question, your prompt enters one of these server rooms. | 当你问LLM一个问题时,你的提示进入了这些服务器的一个房间. |
| 15 | It's too complicated to explore how that model responds right now, | 这太复杂了 无法探索这个模型现在的反应, |
| 16 | but it involves doing tons of math. | 但它涉及做吨的数学。 |
| 17 | To manage these calculations, a data center relies on thousands of server racks, | 为了管理这些计算,一个数据中心依赖于数千个服务器架, |
| 18 | each stacked with trays containing server boards that house memory chips and GPUs. | 每个都堆放着包含内存芯片和GPU的服务器板的托盘. |
| 19 | These powerhouse microchips use thousands of tiny computing cores, | 这些电源芯片使用了数千个微小的计算核心, |
| 20 | working in parallel to do trillions of mathematical operations per second. | 平行工作,每秒完成数以万亿计的数学操作. |
| 21 | While most prompts only require a few GPUs to answer, | 虽然大多数提示只需要几个GPU来回答, |
| 22 | data centers already field tens of thousands of prompts a second, | 数以万计的数据中心 已经出现一秒 |
| 23 | and that rate is growing fast. | 而这个速度正在迅速增长。 |
| 24 | Plus, training LLMs requires hundreds of thousands of GPUs working together, | 此外,培训LLMS需要数十万GPU合作, |
| 25 | making server grids like these essential to get an AI up and running. | 制作这样的服务器网格 以获得一个AI运行。 |
| 26 | However, all this energy-intensive work leads to a major problem—heat. | 然而,所有这些能源密集型工作都导致了一个重大问题——热。 |
| 27 | Even in climate-controlled server rooms, GPUs get hot enough to melt server boards, | 即使在气候控制的服务器室, GPU 得到足够热 熔化服务器板, |
| 28 | and the copper wires connecting these chips constantly dissipate energy, | 连接这些芯片的铜线 不断分散能量 |
| 29 | using more electricity and generating additional heat. | 使用更多的电力并产生额外的热量。 |
| 30 | Fortunately, engineers are on the job. | 幸运的是,工程师在工作。 |
| 31 | To cool chips, they're using tried-and-true solutions like water towers | 为了冷却芯片,他们正在使用像水塔这样的 试探和真实的解决方案 |
| 32 | that absorb and evaporate heat, alongside new approaches like piping chilled water | 吸收和蒸发热量 与新的方法,如管道冷水 |
| 33 | behind server boards or dunking entire trays into non-conductive cooling liquids. | 在服务器板后面或者将整个托盘灌入非导冷却液中. |
| 34 | At the same time, they're replacing inefficient copper wires | 同时,他们取代了低效的铜线 |
| 35 | that transmit data through electricity with optical fibers that transmit data through light. | 用光纤通过电传输数据,光纤通过光传输数据。 |
| 36 | At the top of many racks, optical transceivers convert electrical data to light signals, | 在很多机架中,光学收发器将电气数据转换为光信号, |
| 37 | which travel through optical fibers with higher data fidelity than copper and near zero power loss. | 它通过光纤 数据忠实度高于铜 几乎零功率损失。 |
| 38 | And as photonics technology gets further integrated inside servers, | 随着光子科技进一步融入服务器 |
| 39 | it can solve another major problem. | 它可以解决另一个重大问题。 |
| 40 | GPUs only have so much space for copper connections, | GPU只有这么多的铜连接空间, |
| 41 | limiting how much data they transmit at once. | 限制它们同时传输的数据数量。 |
| 42 | But integrated photonic chips can use different wavelengths of light | 但集成光子芯片可以使用不同波长的光 |
| 43 | to send multiple streams of data over a few optical lines. | 以在几条光学线上发送多个数据流。 |
| 44 | 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GPU在邻近的托盘和架子上与内存芯片相互作用, |
| 45 | spreading out their heat load and allowing bigger LLMs to use less power. | 分散他们的热负荷 并允许更大的LLMs使用更少的动力。 |
| 46 | While some experts believe integrated photonics will reduce data center power use by nearly 50%, | 虽然一些专家认为综合光子能将数据中心的功率减少近50%, |
| 47 | this still might not be enough. | 这也许还不够。 |
| 48 | GPUs run at roughly 16 light bulbs worth of energy. | GPU运行 大约16灯泡 价值能量。 |
| 49 | Multiply that by every GPU in a data center, | 乘以数据中心的每个GPU, |
| 50 | and you're looking at enough energy to power 300,000 homes. | 你正在寻找足够的能量 为30万个家庭提供动力。 |
| 51 | Companies are also pursuing ever-smarter AI, | 公司也在追求更聪明的AI, |
| 52 | which so far has meant ever-larger AI. | 至今为止,它意味着更大的AI。 |
| 53 | Even with increasing efficiencies, larger models require bigger data centers and more electricity. | 即使提高效率,更大的模型也需要更大的数据中心和更多的电力。 |
| 54 | Some estimates predict that by 2028, | 一些估计预测到2028年, |
| 55 | U.S. data centers alone could be using as much energy as eight New York cities. | 光是美国数据中心 就能用到8个纽约城市的能量 |
| 56 | In fact, some AI companies have already proposed pairing next-gen data centers | 事实上,一些AI公司已经提议 配对下一代数据中心 |
| 57 | with their own personal nuclear power plant. | 拥有自己的个人核电站 |
| 58 | Clearly, we don't have enough energy to support this rapid expansion, | 显然,我们没有足够的能量 支持这个快速扩张, |
| 59 | so we either need to hit the brakes or find ways to compute smarter. | 所以我们要么要踩刹车 要么想办法更聪明地计算 |
| 60 | For example, while current LLMs are designed to answer every possible question, | 例如,虽然目前的LLMs旨在回答每一个可能的问题, |
| 61 | you don't need a model that knows Shakespeare to help you fix your dishwasher. | 你不需要一个知道莎士比亚的模特 帮你修洗碗机 |
| 62 | We could design topic-specific small language models stored in micro data centers. | 我们可以设计出存储在微数据中心的针对特定主题的小语言模型。 |
| 63 | Alternately, tomorrow's LLMs might employ new forms of optical computing | 换句话说,明天的LLMS可能会采用新的光学计算方式 |
| 64 | that cut energy costs by using light itself to do the math. | 利用光本身做数学来降低能量成本 |
| 65 | But it won't matter how efficient these systems get if our hunger for AI keeps increasing. | 但是,如果我们对AI的渴望不断增加,这些系统的效率会如何,这不重要. |
| 66 | Ultimately, this isn't a technical problem. | 归根结底,这不是一个技术问题。 |
| 67 | It's a social one. | 社'. |
| 68 | Just how many resources are we willing to spend to build smarter machines? | 我们愿意花多少资源来制造更聪明的机器? |
| 69 | Microchips don't just power data centers. | 微芯片不只是电源数据中心。 |
| 70 | They power every electronic device we use today, | 他们为我们今天使用的每一个电子设备提供动力, |
| 71 | from phones, cars, appliances, medical tools, and many others. | 电话、汽车、电器、医疗工具等等 |
| 72 | Check out this video to find out how they're manufactured | 看看这段视频,看看他们是如何制造的 |
| 73 | and explore what can be done about their environmental impact. | 探索如何应对其环境影响。 |
该视频共有字幕 73 条。解锁更多字幕为会员功能，请移动到 价格