Satya Nadella – How Microsoft thinks about AGI

章节 1：Fairwater 2 数据中心：AI 算力的十倍扩张

📝 本节摘要：

Satya Nadella（萨提亚·纳德拉）和 Scott Guthrie 接受了采访，并带领访谈者参观了微软全新的 Fairwater 2 数据中心，该中心是目前世界上最强大的设施。微软致力于每 18 到 24 个月将训练容量提升 10 倍。Fairwater 2 的网络光纤数量几乎相当于两年前 Azure 所有数据中心的总和，旨在为未来的大型训练任务聚合计算能力（flops），并能够在两个不同区域进行训练，预计实现比训练 GPT-5 时高出十倍的增长。

[原文] Today we are interviewing Satya Nadella.

[译文] 今天我们正在采访萨提亚·纳德拉（Satya Nadella）。

[原文] "We" being me and Dylan Patel, who is founder of SemiAnalysis.

[译文] “我们”是指我，以及 SemiAnalysis 的创始人迪伦·帕特尔（Dylan Patel）。

[原文] Satya, welcome.

[译文] 萨提亚，欢迎。

[原文] Thank you. It's great. Thanks for coming over to Atlanta.

[译文] [SN]: 谢谢。太棒了。谢谢你们来到亚特兰大。

[原文] Thank you for giving us the tour of the new facility.

[译文] 谢谢您带我们参观新设施。

[原文] It's been really cool to see.

[译文] 看到它真是太酷了。

[原文] Absolutely.

[译文] [SN]: 绝对的。

[原文] Satya and Scott Guthrie, Microsoft's EVP of Cloud and AI, give us a tour of their brand new Fairwater 2 data center, the current most powerful in the world.

[译文] 萨提亚和微软云与人工智能执行副总裁（EVP of Cloud and AI）斯科特·格斯里（Scott Guthrie）带我们参观了他们全新的 Fairwater 2 数据中心，这是目前世界上最强大的数据中心。

[原文] We've tried to 10x the training capacity every 18 to 24 months.

[译文] 我们一直努力在每 18 到 24 个月内将训练容量增加 10 倍。

[原文] So this would effectively be a 10x increase from what GPT-5 was trained with.

[译文] [Q]: 那么，这实际上将比训练 GPT-5 时所用的容量增加 10 倍。

[原文] So to put it in perspective in the number of optics, the network optics in this building is almost as much as all of Azure across all our data centers two and a half years ago.

[译文] [SN]: 所以从光纤（optics）数量的角度来看，这座建筑中的网络光纤几乎相当于我们两年前所有 Azure 数据中心（data centers）的总和。

[原文] It's got like five million network connections.

[译文] 它有大约五百万个网络连接。

[原文] You've got all this bandwidth between different sites in a region and between the two regions.

[译文] 你在同一区域内不同站点之间，以及在两个区域之间，都拥有所有这些带宽。

[原文] So is this like a big bet on scaling in the future, that you anticipate in the future that there's going to be some huge model that will

[译文] 那么，这是否是对未来扩展（scaling）的一大押注，即您预期未来会出现某个需要两个完全不同的区域来训练的巨大模型?

[原文] require two whole different regions to train?

[译文] 需要两个完全不同的区域来训练?

[原文] The goal is to be able to aggregate these flops for a large training job and then put these things together across sites.

[译文] [SN]: 目标是能够为大型训练任务聚合这些浮点运算能力（flops），然后将这些资源跨站点整合起来。

[原文] The reality is you'll use it for training and then you'll use it for data gen, you'll use it for inference in all sorts of ways.

[译文] 现实情况是，你将用它进行训练，然后用它进行数据生成（data gen），你将以各种方式用它进行推理（inference）。

[原文] It's not like it's going to be used only for one workload forever.

[译文] 它不会永远只用于一种工作负载（workload）。

[原文] Fairwater 4, which you're going to see under construction nearby, will also be on that one petabit network so that we can actually

[译文] 你们即将看到正在附近建设的 Fairwater 4，也将接入该一拍比特（petabit）网络，这样我们就可以真正地

[原文] link the two at a very high rate.

[译文] 以非常高的速率连接这两个设施。

[原文] Then we do the AI WAN connecting to Milwaukee where we have multiple other Fairwaters being built.

[译文] 接下来，我们进行人工智能广域网（AI WAN），连接到正在建设多个 Fairwater 设施的密尔沃基（Milwaukee）。

[原文] Literally you can see the model parallelism and the data parallelism.

[译文] 实际上你可以看到模型并行（model parallelism）和数据并行（data parallelism）。

[原文] It's kind of built for, essentially, the training jobs, the super pods across this campus.

[译文] 它本质上是为训练任务，即跨越这个园区的超级计算集群（super pods）而构建的。

[原文] And then with the WAN, you can go to the Wisconsin data center.

[译文] 然后通过广域网（WAN），你可以连接到威斯康星州（Wisconsin）的数据中心。

[原文] You literally run a training job with all of them getting aggregated.

[译文] 你确实可以运行一个聚合了所有这些资源进行训练的任务。

[原文] What we're seeing right here is a cell with no servers in it yet, no racks.

[译文] 我们现在看到的这里是一个单元（cell），里面还没有服务器，也没有机架（racks）。