Name: 為什麼馬斯克要建造世界上最大的超級電腦？ (Why Elon Musk Is Building The World's BIGGEST Supercomputer)
Uploaded: 2024-07-08T07:00:14.000Z
Duration: 15 min 52 s

反身代名詞

Elon Musk appears to have found himself a new obsession, supercomputers.

埃隆-馬斯克（Elon Musk）似乎又找到了一個新的痴迷對象--超級計算機。

Elon has not been shy about telling the world that he is laying down fat stacks of cash to build out his computer army.

埃隆毫不避諱地告訴全世界，他正在投入大筆資金，打造他的電腦軍團。

We are looking at tens of billions of dollars being invested in just the next year alone.

僅在明年，我們就將看到數百億美元的投資。

Because well, that's where things get interesting.

因為這就是事情變得有趣的地方。

Okay first let's break this whole supercomputer situation up into some manageable chunks and start to identify what goes where because Elon is building multiple computer clusters for different projects that do different things and at some point he seemingly referred to each of them as either the world's biggest or the world's most powerful so it's a bit hard to follow.

好吧，首先讓我們把整個超級計算機的情況抽成幾個易於管理的小塊，並開始確定什麼是什麼，因為埃隆正在為不同的項目建造多個計算機集群，做不同的事情，在某些時候，他似乎把每個集群都稱為世界上最大的或世界上最強大的，所以這有點難以理解。

They're getting a brand new supercomputer.

他們將獲得一臺全新的超級計算機。

This is actually under construction right now.

The Gigafactory in Austin, Texas, which was already the biggest car manufacturing plant in the world, is getting even bigger and most of that new real estate is being dedicated to supercomputer activity, which is why they are also building this whole dedicated cooling bunker.

位於德克薩斯州奧斯汀的 Gigafactory 工廠已經是世界上最大的汽車製造廠，現在它的規模變得更大，其中大部分新廠房都將用於超級計算機活動，這就是為什麼他們還要建造這一整座專用冷卻掩體的原因。

埃隆最近在 X 上談到了這一點。

He wrote, sizing for about 130 megawatts of power and cooling this year but will increase to over 500 megawatts over the next 18 months or so.

他寫道，今年的電力和製冷規模約為 130 兆瓦，但在未來 18 個月左右將增至 500 兆瓦以上。

So as for what exactly Tesla plans to do with all that power, we'll get into those details shortly but first let's talk about an even bigger supercomputer project that Elon is working on for his XAI venture.

至於特斯拉計劃如何利用這些能量，我們稍後會詳細介紹，但首先讓我們來談談埃隆正在為他的 XAI 風險投資公司開發的一個更大的超級計算機項目。

This one will be built in Memphis, Tennessee and Elon has referred to it as the Gigafactory of Compute.

這座工廠將建在田納西州的孟菲斯，埃隆將其稱為 "計算的 Gigafactory"。

He then went on to tell officials, quote, my vision is to build the world's largest and most powerful supercomputer and I'm willing to put it in Memphis.

他接著對官員們說，我的願景是建造世界上最大、最強大的超級計算機，我願意把它放在孟菲斯。

And then in addition to all of that, Tesla has been kind of secretly converting all of their vehicles into a mobile, decentralized supercomputer network of their own, which will eventually include all of their humanoid robots as well.

除此以外，特斯拉還一直在祕密地將他們的所有車輛轉換成一個移動的、分散的超級計算機網絡，最終還將包括他們所有的仿人機器人。

Okay, let's start breaking down how all of this stuff works.

好了，讓我們開始分析這些東西是如何工作的。

So as many of you know, we are all Canadians here behind the scenes, which is generally pretty awesome except when it comes to the internet.

你們很多人都知道，我們都是幕後的加拿大人，除了在互聯網上，一般情況下都很了不起。

The Canadian government has a weird thing about trying to control what we can and can't see online.

加拿大政府有一個怪癖，就是試圖控制我們在網上能看到什麼，不能看到什麼。

It's even started to result in Meta and Google limiting the type of content that I'm able to access.

這甚至導致 Meta 和 Google 開始限制我能夠訪問的內容類型。

And if that wasn't enough, our Netflix selection sucks up here, too.

如果這還不夠，我們的 Netflix 選擇在這裡也很糟糕。

So in an effort to dodge all of that madness, I've started using a VPN.

是以，為了躲避這些瘋狂行為，我開始使用 VPN。

More specifically, I've been using CyberGhost VPN, not only because they have a cool name, but also because they offer a very high quality product at an even more affordable price and a cool name.

更具體地說，我一直在使用 CyberGhost VPN，不僅因為他們有一個很酷的名字，還因為他們以更實惠的價格和很酷的名字提供了非常優質的產品。

With CyberGhost VPN, all of your traffic goes through a secure VPN tunnel.

使用 CyberGhost VPN，您的所有流量都將通過安全的 VPN 隧道傳輸。

Your IP address is hidden and your data is encrypted.

您的 IP 地址會被隱藏，您的數據也會被加密。

Trust me, you want your data and history secured and encrypted.

相信我，你希望你的數據和歷史記錄得到安全加密。

So what you do online is strictly your business.

是以，你在網上做什麼完全是你自己的事。

You can also change your online location in just three clicks and get access to geo-restricted content from dozens of streaming services like Netflix.

您還可以只需點擊三下就能更改您的在線位置，並從 Netflix 等數十種流媒體服務獲取受地理位置限制的內容。

With CyberGhost VPN, you can even find better online shopping deals or play games blocked in your region.

使用 CyberGhost VPN，您甚至可以找到更好的在線購物優惠，或者玩您所在地區封鎖的遊戲。

CyberGhost VPN is available for all platforms such as Windows, Mac OS, Android, iOS, and many others.

CyberGhost VPN 適用於 Windows、Mac OS、Android、iOS 等所有平臺。

You can use one subscription to protect up to seven devices at a time, so you can easily share with your family and friends.

一次訂購最多可保護七臺設備，方便您與家人和朋友共享。

Plus, you get a 45-day money-back guarantee and 24-7 customer support, so everything is risk-free.

此外，您還將獲得 45 天退款保證和全天候客戶支持，是以一切都沒有風險。

CyberGhost VPN is offering their best deal ever, just $2.03 per month, and you get four months free, which is 84% off.

CyberGhost VPN 提供了有史以來最優惠的價格，每月僅需 2.03 美元，還可免費使用四個月，折扣高達 84%。

So, join over 38 million happy users, including me, and click the link in our description to sign up today.

是以，請加入包括我在內的 3800 多萬快樂用戶的行列，點擊說明中的鏈接立即註冊。

Okay, that brand-new GigaTexas data center that's under construction right now is going to be aimed specifically at developing Tesla's full self-driving software.

好吧，正在建設中的全新 GigaTexas 數據中心將專門用於開發特斯拉的全套自動駕駛軟件。

And to do that, Elon says that it will be loaded with a combination of NVIDIA GPUs along with Tesla's own AI-specific hardware.

埃隆表示，為了實現這一目標，特斯拉將結合英偉達™（NVIDIA®）圖形處理器和特斯拉自己的人工智能專用硬件。

Elon says that the south extension of the factory is custom-built for heavy power compute and cooling, hence the giant cooling bunker.

埃隆說，工廠南面的擴建部分是為大功率計算和冷卻而定製的，是以有一個巨大的冷卻掩體。

He said earlier this month that the initial system is 50,000 NVIDIA H100 GPUs alongside 20,000 Tesla hardware for AI computers and a massive network of hard drives for video storage.

他在本月早些時候說，最初的系統由 5 萬個英偉達 H100 GPU 和用於人工智能計算機的 2 萬個特斯拉硬件以及用於視頻存儲的大規模硬盤網絡組成。

When Elon says that Tesla is spending $10 billion this year on AI, this is where a large chunk of that cash is going right now.

當埃隆說特斯拉今年將在人工智能上投入 100 億美元時，這就是其中一大筆現金的去向。

The NVIDIA H100 is a current industry standard chip for AI training, which is an extraordinarily complicated process involving endless calculations in an attempt to teach a computer network the difference between right and wrong.

英偉達™（NVIDIA®）H100 是目前用於人工智能訓練的行業標準芯片，這是一個異常複雜的過程，涉及無休止的計算，試圖教會計算機網絡區分對與錯。

We provide the input and then train the network until it produces the desired output.

我們提供輸入，然後訓練網絡，直到它產生所需的輸出。

In the most common AI models that we use today, like ChatGPT or MidJourney, we input words or pictures and receive the same as an output.

在我們今天使用的最常見的人工智能模型中，如 ChatGPT 或 MidJourney，我們輸入文字或圖片，然後接收相同的輸出。

In some more advanced new models, you can even input text and get video as an output.

在一些更先進的新型號中，您甚至可以輸入文本並輸出視頻。

In Tesla's AI models, the input is video and the output is driving a car.

在特斯拉的人工智能模型中，輸入是視頻，輸出是駕駛汽車。

So this is significantly more complex than what most people have grown accustomed to, and as a result, it requires a much more powerful and sophisticated hardware setup to make it all work.

是以，這比大多數人所習慣的要複雜得多，也是以需要更強大、更復雜的硬件設置來實現這一切。

That's where Tesla's AI hardware comes into play.

這就是特斯拉人工智能硬件發揮作用的地方。

In addition to the training computer, Tesla also uses a powerful computer inside their vehicles.

除了培訓電腦，特斯拉還在車內使用了功能強大的電腦。

This is called Inference Compute, and it allows the AI to operate in the real world in real time and make decisions at incredibly high speed.

這就是所謂的 "推理計算"，它能讓人工智能在現實世界中實時運行，並以難以置信的高速做出決策。

You know when you ask ChatGPT a question and it has to kind of think for a second before generating an answer?

你知道當你向 ChatGPT 提問時，它需要思考一秒鐘才能給出答案嗎？

That's because pretty much every AI model that we use right now is just a cloud-based web app.

這是因為我們現在使用的幾乎所有人工智能模型都只是一個基於雲的網絡應用程序。

ChatGPT isn't operating on your computer or your phone.

ChatGPT 無法在您的電腦或手機上運行。

The prompt that you write is sent off to an OpenAI data center, where it's processed through their own giant supercomputer cluster, and then the response is sent back to you.

您寫下的提示會被髮送到 OpenAI 數據中心，在那裡通過他們自己的巨型超級計算機集群進行處理，然後再將回復發送給您。

This is fine if you just want a cookie recipe or a picture of a robot on Mars, but when AI is driving your car, it cannot be sending every decision back and forth through the internet.

如果你只是想要一份餅乾食譜或一張火星上機器人的照片，這樣做沒問題，但當人工智能在駕駛你的汽車時，它不能通過互聯網來回發送每一個決定。

It needs a brain that will do the thinking and decision-making on-site.

它需要一個能現場進行思考和決策的大腦。

Not only is this inference hardware inside every car, it's also used in the Tesla data centers, alongside the NVIDIA hardware.

這種推理硬件不僅安裝在每輛車內，還與英偉達（NVIDIA）硬件一起用於特斯拉數據中心。

The way that Elon explains this is that the H100s, the Tesla hardware and software video data, is all part of one big training loop.

埃隆對此的解釋是，H100s、特斯拉硬件和軟件視頻數據都是一個大訓練循環的一部分。

Currently, Tesla is using their AI hardware version 4, but Elon says that before the end of next year, there will be hardware version 5, which is called AI 5.

目前，特斯拉使用的是第 4 版人工智能硬件，但埃隆表示，在明年年底之前，將推出第 5 版硬件，即人工智能 5。

This will roll out in vehicles and data centers.

Elon says that the new chip will have 10 times more capability than hardware 4, so that would potentially equal 10 times better performance in the training loop and 10 times better performance running the AI model in the vehicle for autonomous driving.

埃隆說，新芯片的性能將是硬件 4 的 10 倍，是以在訓練環中的性能可能會提高 10 倍，在車輛中運行人工智能模型以實現自動駕駛的性能也會提高 10 倍。

The same AI5 chip is not only going into future Tesla vehicles, it's also going into Tesla's This is likely the motivating factor for Tesla building out such a massive new supercomputer cluster.

同樣的 AI5 芯片不僅將應用於未來的特斯拉汽車，還將應用於特斯拉的計算機系統。 這很可能是特斯拉建立如此龐大的新超級計算機集群的動因。

Obviously, full self-driving still needs some work, but Elon said a couple of months back that Tesla was not compute-constrained for FSD.

顯然，完全自動駕駛還需要一些工作，但埃隆幾個月前曾說過，特斯拉在 FSD 方面的計算能力並不受限。

Though when it comes to developing brand new models for the Tesla bot, those could prove to be even bigger than full self-driving.

不過，在為特斯拉機器人開發全新車型時，這些車型可能會比完全自動駕駛更重要。

There is so much that the bot will need to know in order to function in the real world and perform human tasks, so Tesla will not be done with AI training any time soon, even if we do get full robotaxis next year.

為了在現實世界中發揮作用並執行人類任務，機器人需要了解的東西太多了，是以特斯拉不會在短期內完成人工智能培訓，即使我們明年就能獲得完整的機器人軸。

And if that wasn't enough, Elon already has some big ideas for what Tesla could do with all of these high-powered inference computers out there in the world.

如果這還不夠的話，埃隆已經有了一些偉大的想法，他想讓特斯拉利用世界上所有這些強大的推理計算機做些什麼。

At the recent Tesla shareholder meeting in June, Elon Musk put out this concept about how the company might be able to use future autonomous vehicle hardware in some unconventional ways.

在最近於 6 月份舉行的特斯拉股東大會上，埃隆-馬斯克提出了這樣一個概念，即公司可能會以一些非常規的方式使用未來的自動駕駛汽車硬件。

The idea is that if there is a point down the road where tens of millions of Tesla vehicles are out there in the world, and many of them are equipped with this AI5 hardware, then you essentially have a giant, decentralized supercomputer made up of individual cars that are all connected by the Tesla network.

我們的想法是，如果有一天，全世界有數以千萬計的特斯拉汽車，其中許多都配備了 AI5 硬件，那麼你就擁有了一個由特斯拉網絡連接的單個汽車組成的巨大、分散的超級計算機。

Elon speculated that by around the end of the decade, there might be a gigawatt of computing power when all of these Tesla vehicles are combined.

埃隆推測，到本世紀末，當所有這些特斯拉汽車加在一起，計算能力可能會達到千兆瓦。

Going back to that GigaTexas computer cluster, there's up to 500 megawatts of power, which is half a gigawatt.

回溯到 GigaTexas 計算機集群，其功率高達 500 兆瓦，相當於半個千兆瓦。

Anyway, the in-car computer of a robotaxi would be occupied with self-driving tasks most of the time, but for periods when the vehicle is either recharging or just not being used, that computing power might be harnessed to do other important jobs.

總之，機器人計程車的車載電腦大部分時間都在執行自動駕駛任務，但在車輛充電或閒置期間，這些計算能力可能會被用來執行其他重要任務。

This is basically the same concept behind Amazon Web Services.

這與亞馬遜網絡服務背後的概念基本相同。

This is the business model that actually makes money at Amazon, and the concept was born when the company found that all of the data processing hardware that they required to deal with peak traffic situations like Christmas was just sitting around and not being utilized during the other times of the year.

這就是亞馬遜真正賺錢的商業模式，當該公司發現他們為應對聖誕節等流量高峰而需要的所有數據處理硬件在一年中的其他時間都閒置不用時，這一概念就應運而生了。

So they started renting out that computer hardware to other companies that needed the processing but didn't have the capital to build their own data center.

於是，他們開始將計算機硬件出租給其他需要處理能力但沒有資金自建數據中心的公司。

In theory, Tesla's Robotaxi network could kind of do the same thing, even the consumer vehicles as well.

從理論上講，特斯拉的 Robotaxi 網絡也可以做同樣的事情，甚至消費車輛也可以。

You could check a box on the app that would let Tesla access your in-car computer and maybe you get a cut of the revenue it earns.

您可以在應用程序上勾選一個選項，讓特斯拉訪問您的車載電腦，或許您還能從特斯拉賺取的收入中分一杯羹。

It's kind of like a crypto mining pool if you've ever seen or tried one of those.

這有點像加密貨幣礦池，如果你見過或嘗試過這種礦池的話。

Okay, so we've talked about AI in cars and autonomous driving, that's all cool, but what about solving the mysteries of the universe?

好了，我們已經談到了汽車和自動駕駛中的人工智能，這一切都很酷，但如果要解開宇宙之謎呢？

Okay, let's go back to Musk's Gigafactory of Compute, which is not to be confused with his giant computer at the Tesla Gigafactory.

好吧，讓我們回到馬斯克的 "計算千兆工廠"（Gigafactory of Compute），不要把它與他在特斯拉千兆工廠（Tesla Gigafactory）的巨型計算機混為一談。

This is the new installation in Memphis that will be dedicated to XAI.

這是在孟菲斯專門為 XAI 安裝的新設施。

This is the one that could potentially become the world's largest and most powerful supercomputer by the end of 2025 if you're inclined to follow Elon timelines.

如果你願意追隨埃隆的時間表，這臺超級計算機有可能在 2025 年底成為世界上最大、最強大的超級計算機。

In the short term, Elon is targeting 100,000 of the NVIDIA H100 GPUs up and running before the end of this year.

在短期內，埃隆的目標是在今年年底前安裝並運行 10 萬臺英偉達™（NVIDIA®）H100 GPU。

That's what he believes the company needs to build out the next generation of their Grok AI language model.

他認為，這正是公司建立下一代 Grok 人工智能語言模型所需要的。

This is chatbot that is native to the X platform, available to premium subscribers.

這是 X 平臺原生的哈拉機器人，高級用戶可以使用。

It's like ChatGPT except Grok has real-time access to every post on X and Grok also has the freedom of language to write swear words and make cringe attempts at humor, which can be a lot of fun.

它就像 ChatGPT，只不過 Grok 可以實時查看 X 上的每一條帖子，而且 Grok 還可以自由使用語言，寫髒話和嘗試幽默，非常有趣。

This is just basic level stuff though, XAI was able to build out Grok 1 at an incredibly fast pace, it only took about 6 months from the company being founded to the release of their first product.

這還只是基礎層面的東西，XAI能夠以令人難以置信的速度開發出Grok 1，從公司成立到發佈第一款產品只用了大約6個月時間。

Although it still lags behind ChatGPT a little bit in terms of capability and Grok definitely lacks the name recognition and popularity currently being enjoyed by ChatGPT and OpenAI.

雖然在功能方面，Grok 仍略遜於 ChatGPT，而且在知名度和受歡迎程度方面，Grok 肯定不如 ChatGPT 和 OpenAI。

Grok 2 might be able to start turning the tide.

This is the product that XAI is working on currently.

這就是 XAI 目前正在開發的產品。

The upgrade will allow Grok to both interpret and produce images and visual media so you can have it turn a spreadsheet into a graph or turn a graph into a spreadsheet or identify the content of a photograph or even explain a piece of art or a meme.

升級後，Grok 將能夠解釋和製作影像和視覺媒體，是以您可以讓它將電子表格轉換成圖表，或將圖表轉換成電子表格，或識別照片內容，甚至解釋一件藝術品或備忘錄。

So far, XAI has been able to do all of this AI training in partnership with a company called Oracle Cloud.

到目前為止，XAI 已經能夠與一家名為 Oracle Cloud 的公司合作完成所有這些人工智能培訓。

Just like we were talking about earlier with renting out data processing, that's exactly what XAI has done here.

就像我們之前談到的出租數據處理一樣，這正是XAI在這裡所做的。

It looks like they have been renting the equivalent of about 20,000 H100 GPUs from Oracle.

看起來他們從甲骨文公司租用了相當於約 20,000 個 H100 GPU 的設備。

Elon has said that the 100,000 H100 cluster will be necessary to train Grok 3, which is still unknown but we have to assume that this would be moving in the direction of a generalized artificial intelligence, something that can deal with text, sound, images, and video as both input and output media, which probably not by coincidence would be the exact kind of AI model that the Tesla Bot will need to fully function as a productive member of society at some point in the future.

埃隆曾說過，訓練 Grok 3 需要 10 萬個 H100 集群，目前還不得而知，但我們必須假設，這將朝著通用人工智能的方向發展，能夠處理文本、聲音、影像和視頻等輸入和輸出媒體。

Now beyond that is where things start to get weird.

除此之外，事情就開始變得奇怪了。

Elon's vision of his completed Gigafactory of Compute is now looking like 300,000 units of the B200 GPU, which is Nvidia's next big chip release.

埃隆對建成的 "計算千兆工廠 "的願景，現在看起來就像 30 萬臺 B200 GPU，這是 Nvidia 發佈的下一款大型芯片。