人工智慧加速一個國家：丹麥人工智慧超級電腦 Gefion 的進步：歐洲、中東和非洲地區問答 [S72318a]

Nadia Carlsten,丹麥人工智慧創新中心首席執行官

了解一項令人興奮的策略合作，它為丹麥帶來了大規模人工智慧運算能力，以及新的人工智慧超級電腦如何加速許多創新者的工作。丹麥人工智慧創新中心 (DCAI) 的目標是加速醫療保健、生命科學到綠色轉型等領域的研究和創新，支持開發解決世界上最大問題的創新解決方案。聽聽 DCAI 執行長 Nadia Carlsten 介紹如何建立世界上最大的超級電腦之一，以及來自學術界、新創公司和大公司的客戶如何利用這個新設施在醫療保健、生命科學和量子計算領域取得進展。
* 丹麥在人工智慧方面投入巨資，與 NVIDIA 合作將超級電腦引入丹麥
* 由 NVIDIA 提供支援的新型 AI 超級電腦是加速丹麥研究的重要工具，包括醫療保健、生命科學、能源和量子計算等領域的研究
* 為了充分利用新的超級計算機，重要的是確定正確的用例，以充分利用 NVIDIA 的規模和架構

主題：模擬/建模/設計 - 量子計算
行業細分：所有行業
技術等級：商務/行政
目標受眾：企業主管
所有目標受眾類型：企業主管、研究：學術、研究：非學術
3 月 21 日，星期五
晚上 8:00 - 晚上 8:50 中部標準時間

AI逐字稿
謝謝大家今天來參加這個會議。嗯，所以我們是 AGTC，而 GTC 的核心就是 AI 卓越（AI Excellence）。

Thank you everybody for coming to this session today. Um, so we are AGTC, and GTC is all about AI Excellence.

我們可能都聽過很多關於 AI 如何加速各種任務的會議，比如 AI 加速工作流程（workflows）、AI 加速整個產業（industries），無論是製藥（pharma）還是工業自動化（industrial automation）。今天我要跟大家談的是 AI 如何加速整個國家，這可是一項艱鉅的任務，對吧？也是一個很重要的議題。

We’ve probably all heard a lot of sessions about AI accelerating different tasks, like AI accelerating workflows, AI accelerating entire industries, whether it’s pharma or industrial automation. What I’m going to talk to you about today is AI accelerating an entire country, which is quite a task, right? Quite a session.

我的名字不是問號，我是丹麥 AI 創新中心（Danish Center for AI Innovation）的首席執行官。今天我要跟大家分享的是我們的全新 AI 超級電腦（AI supercomputer）的故事。這台超級電腦幾乎是全新的，至少按照我們的時間表來說是這樣，可能在 AI 的時間表裡不算什麼，但它是在去年十月剛剛推出的。我會告訴你們我們為什麼決定建造它，這段旅程是怎麼進行的，更具體來說，還會講到我們的客戶如何使用它，因為這一點真的很重要。

My name is not a question—I am the CEO of the Danish Center for AI Innovation. And today, I’m going to talk to you about our brand-new AI supercomputer. It’s pretty much brand new, at least by our timelines—maybe not by AI timelines—but it just launched in October. I’m going to tell you the story about why we decided to build it, how that journey has been, and more specifically, how our customers are using it, because that is really important.

那麼，這段旅程是怎麼開始的呢？就像很多與 AI 相關的事情一樣，它始於 GTC，而且是從去年才開始的。當時有一個來自丹麥的代表團來到 GTC，特別是為了宣布一個重大消息，那就是 Nvidia 與諾和諾德基金會（Novo Nordisk Foundation）合作。他們計劃聯手打造一台巨大的 AI 電腦，這是丹麥有史以來建造過的最大規模的設備。當時他們有一個宏大的願景，那是去年的事了，但那時還沒有 GPU，甚至連一台電腦都沒有，只是一個非常有雄心的願景和合作的協議。

So, how did this journey start? Like many things with AI, it started at GTC, and it only started last year. There was a delegation from Denmark that came to GTC, specifically to make this big announcement: Nvidia had partnered with the Novo Nordisk Foundation. They were going to team up to build this massive AI computer—something bigger than had ever been built in Denmark before. They had this big vision a year ago, but there were still no GPUs, let alone a computer. It was just a very ambitious vision and agreements to work together.

這就是事情的進展情況。到了去年十月的啟動儀式上，我們很榮幸邀請到了 Jensen（Nvidia 的 Jensen Huang）來到哥本哈根參加我們的啟動儀式。如果你不認識另一位參與者，那其實是丹麥國王，這一點就足以說明這個項目對丹麥的重要性。對丹麥來說，這是一件大事，對整個 AI 社群來說也是如此。這真的是件非常重要的事情。我們非常感謝 Nvidia 對這個重大項目的支持，他們一直以來都非常配合。

And this is how it’s been going. At our launch in October, we had the pleasure of having Jensen come to our launch in Copenhagen. In case you don’t recognize the other guy, that’s actually the King of Denmark, and that just goes to show the importance of this project for Denmark. It’s a huge thing for Denmark, but also for the whole AI community. It’s been a really, really big thing. And we’re so grateful that it’s also been a really important thing for Nvidia and that they’ve been so supportive of this huge project.

我們所做的是打造了一個 AI 工廠（AI factory），而且我們在創紀錄的時間內完成了這一切。這就是我想跟大家分享的故事。我們可以把這個故事分成四個部分：首先是規劃階段（planning phase），這也包括成立丹麥 AI 創新中心，這是非常關鍵的一步，因為總得有人來擁有並運營這台電腦，而那就是我們。

What we’ve done is built an AI factory, and we’ve done it in record time. That’s the story I want to tell you. We can basically divide that story into four parts: starting with the planning phase, which also included establishing the Danish Center for AI Innovation, which was really the key part because somebody needs to own and run this computer—and that’s us.

我還會談到建造過程（building phase），這是很有趣的部分，包括組裝電腦並確保它能正常運作。然後是更有趣的部分，也就是讓用戶開始使用這台電腦。這個過程的美妙之處在於，一旦你讓客戶開始使用任何產品，尤其是電腦，他們會開始發現問題，然後你就得重新修復。所以我們和很多客戶一起經歷了一段非常有趣的旅程。

I’m also going to talk about the building phase—that’s the fun part—assembling the computer and making sure it’s working. Then there’s an even more fun part, which is getting users on the computer. The beauty of that is, once you get customers on any product—especially a computer—they start breaking things, and then you have to rebuild again. So, there’s been a really fun journey iterating with a lot of our customers.

我還會提到這個項目在短短一年內已經產生的影響，以及我們希望繼續創造的影響。

And then I’m also going to touch on the impact that we are already making just a year into this whole project and that we hope to continue making.

好的，我們就從成立丹麥 AI 創新中心（Danish Center for AI Innovation）開始講起。你們很多人可能會問，為什麼是丹麥？我出於好奇想問一下，在場有沒有人來自丹麥，或者跟丹麥有關聯？哦，哇，好的，有很多人，這真是太棒了。對於那些不是來自丹麥的人來說，你們可能不知道，丹麥正在發生很多很棒的事情，尤其是在科技和數位領域（tech and digital space）。丹麥在最具創新力的國家排名中名列前茅，位居前十；它在數位競爭力（digital competitiveness）方面排名第四。我認為，丹麥有一種很強大的技術採用文化（tech adoption culture），非常注重通過採用最新技術來加速發展。

So, let’s start with establishing the Danish Center for AI Innovation. A lot of you may be asking, why Denmark? Just out of curiosity, is anybody here from Denmark or related to them? Oh, wow, okay, a lot of you—well, that’s amazing. Uh, so for those of you who are not from Denmark, you may not realize that there’s a lot of amazing things going on in Denmark, especially in the tech and digital space. Uh, it is ranked well in the top ten of the most innovative countries. It’s number four in digital competitiveness. Ah, and generally, I would say there’s a huge culture around tech adoption and just making sure that there’s acceleration by adopting the latest technology.

另一個很有趣且與 AI 高度相關的事情是，AI 的核心在於資料（data），包括資料的品質和獨特性。而丹麥在這方面相當特別。它擁有非常豐富且複雜的公民資料庫（citizen data），尤其是健康資料（health data），這些資料非常有用，因為你可以從中獲得很多洞見（insights），這些是你原本無法得到的。所以，利用這些資料進行 AI 研究有很大的潛力，能產生非常有影響力的成果（impactful results）。

Another really interesting thing about Denmark, which is very relevant for AI, is that AI is all about data and the quality of the data and the uniqueness of the data. And Denmark is fairly unique in that. It has a really big collection of very rich and complex data on its citizens, uh, especially when it comes to health data that can be very useful because you can get a lot of insights. So you wouldn’t otherwise have, uh, so there’s a lot of potential to do AI on that data. And that gets something very impactful.

另外一個優勢是，丹麥有一個非常緊密連結的 AI 創新生態系統（AI innovation ecosystem）。它位於歐洲的中心位置，與其他歐洲國家有很好的關係。所以，丹麥具備很多非常優秀的元素。於是就有了這樣一個願景：他們認為需要在國家層面推動加速計算（accelerated computing），來充分利用這些元素，做出一些很酷的事情。

The other nice thing is it’s a very interconnected AI innovation ecosystem. Uh, it’s also very central in Europe and there’s good relationships with other European countries. So there’s a lot of really, really good elements. And uh, there was this vision that they really need to have accelerated computing on a national scale to do something really cool with all those elements.

還有一點值得一提的是，甚至在我們成立之前，丹麥就已經有一個相當強大的深度科技生態系統（deep tech ecosystem）。

One thing to mention is that even before we existed, there was already quite a strong deep tech ecosystem in Denmark.

丹麥在量子技術（quantum technology）方面有著豐富的歷史，而且那裡有很多研究中心正在進行相關研究。生命科學（life sciences）這一點可能顯而易見，但除此之外，AI 和資料科學（data science）的相關計畫也在進行中。這些只是研究項目，有些還是非常大型的研究項目。但我得提到，在這些領域中也有很多大公司。你們可能知道，嗯，不過我們回顧得不算太遠。製藥業（pharmaceutical industry）裡有很多丹麥公司。所以，對 AI 工廠（AI factory）需求的產生，歸結於四個反覆出現的主題，這些是許多創新者和研究人員不斷提出的問題。

Uh, there’s a lot of quantum—there’s a rich history of quantum in Denmark and there’s a lot of research centers going on there. Uh, life sciences, that might be obvious. Um, and but there’s also AI and data science initiatives going on. And this is just research projects and some very large research projects. But I should mention there’s also a lot of big companies in some of these spaces as well. You probably know, uh, no, we’re not going that far back. Uh, your pharmaceutical—all Danish companies. So the need for an AI factory came down to four recurring themes that a lot of innovators and researchers kept bringing up.

第一個問題是先進的計算硬體（advanced computing hardware）不足。GPU 很難找，品質好的 GPU 更難找，尤其是要大量取得的時候特別困難。第二個問題是 GPU 的取得（access to GPUs）很困難。我說的困難是指價格問題（price issue），對於那些不是大公司、無法負擔高價的單位來說尤其如此。所以學術界（academics）和新創公司（startups）就被排除在外了。此外，還有一個需求是希望能即時存取資源（access in a timely manner）。對於那些正在突破界限、使用最新 AI 技術的研究人員來說，等待幾週才能使用其他國家或歐洲的資源實在太久了。

The first one was that there wasn’t enough advanced computing hardware. It was just hard to find GPUs, and it was hard to find good quality GPUs. Um, and it was especially hard to find them in very large numbers. There was a second issue of access—access to GPUs is difficult. And what I mean by difficult is that there’s an issue of price for anybody that’s not a large company that can afford to pay premium prices. So academics and startups were getting left out. Um, and there was also the desire to be able to access things in a timely manner, and for people doing research that was really pushing the envelope and using the latest in AI, having to wait several weeks to access other national resources and European resources, which was taking too long.

我們試圖解決的另一個問題是所需的支援（support）。開始進行大規模 AI 計算（large-scale AI computing）並做平行處理（parallel processing）並不容易，光是提供 GPU 存取還不夠，還需要大量的用戶支援（user support）。最後，我提到過的資料集（data sets）雖然潛力很大，但對資料主權（data sovereignty）也有很多要求。很多資料不能離開丹麥，所以要想從中獲得分析結果和洞見（analytics and insights），就必須在國內完成所有計算（compute）。

Another thing that we’re trying to solve for is the support needed. Um, so it’s not easy to get started on large-scale AI computing and do parallel processing. And there’s a lot of user support that’s needed beyond just access to the GPUs. And finally, the data sets I mentioned—ah, there’s a lot of potential there, but there’s also a lot of requirements around sovereignty of that data. So a lot of that data cannot leave Denmark. And so in order to get the analytics and the insights that you want out of it, you have to do all of that compute in the country.

就像任何公司一樣，第一個挑戰是找到資金（funding）來支持這個項目。在丹麥的情況下，我們有一個很常見的企業基金會系統（system of enterprise foundations）。我們很幸運有了一個最大的基金會作為我們的投資者，那就是諾和諾德基金會（Novo Nordisk Foundation），它的歷史可以追溯到 1924 年。你可以把它想像成類似比爾蓋茨基金會（Bill Gates Foundation）或英國的歡迎信託（Wellcome Trust）這樣的組織。它有一個慈善使命（philanthropic mission），致力於改善人們的健康（health）和地球的可持續性（sustainability）。他們資助了很多研究和創新（research and innovation）。

So, like any company, one of the first hurdles was finding funding to fund this. Uh, in our case, in Denmark, there’s this system of enterprise foundations that are quite common. And we are lucky to have as our investor one of the biggest foundations, which is the Novo Nordisk Foundation, which dates all the way back to 1924. And uh, you can think about it—it’s very similar to the Bill Gates Foundation, for example, or Wellcome Trust. Um, it has a philanthropic mission to improve people’s health and also the sustainability of the planet. Uh, they fund a lot of research and innovation.

這正好是諾和諾德基金會（Novo Nordisk Foundation）的強項，他們資助的許多項目都能從 AI 中受益。所以這對他們來說非常有意義，完全符合他們的使命（mission）。但這對諾和諾德來說也是一筆非常大的投資。我們還得到了丹麥出口與投資基金（Export and Investment Fund of Denmark）的資助，這家基金通常投資於新創公司（startups）以及投資新創的基金。但當他們認為某個項目能提升整個生態系統（ecosystem）時，也會進行策略性投資（strategic investments）。我們就是其中一個策略性投資項目。總計下來，初始投資約為一億歐元，這只是啟動資金（startup capital），用來購買初始硬體並啟動項目。

So this is right in their wheelhouse, and actually a lot of the projects that they are funding could benefit from AI. So this made a lot of sense for them, lining up very nicely with their mission. Uh, but it’s also a very big investment for Novo Nordisk to be doing. We are also funded by the Export and Investment Fund of Denmark, which typically invests in startups and funds that invest in startups. But they also make strategic investments when they think it’s going to raise up the entire ecosystem. And we’re one of those strategic investments for them. So all combined, the initial investment was about a hundred million euros—that was just the startup capital to get the initial hardware and get going.

我們的獨特之處之一，尤其是因為我們的資金結構（funding structure），在於我們被設定為服務整個生態系統。我們能與學術界（academics）合作，也能與商業領域合作，包括新創公司和大型企業。所以我們採用商業模式（commercial model），每個人都可以使用我們的計算能力（computing capabilities）。但我們有不同的機制，確保有需要的人都能負擔得起費用。我們也與更廣泛的社群合作，這一點也很重要，例如與公共部門機構（public sector institutions）合作，甚至在某些情況下與個人合作，如果他們沒有其他代表且需要使用計算資源（compute）。

One of the things that makes us unique, and is especially relevant because of the funding structure that we have, is we are set up so that we can serve the entire ecosystem. We are able to work with academics, we’re also able to work with businesses—and that’s startups as well as large companies. So we are a commercial model. Uh, everybody can use our computing capabilities and the infrastructure. But then we have different mechanisms to make sure that people can afford to pay if they have that need. We also work with the community at large, which is also really important to be able to work with public sector institutions, for example, and even individuals in some cases, if they’re not otherwise represented and they need to access the compute.

我們的價值主張（value proposition）是提供一站式服務（one-stop shop）的高級計算（advanced computing），特別是能加速研究的先進計算，讓 AI 在不同研究領域產生正面影響（positive impact）。我們提供尖端計算資源（cutting-edge compute），這是我們現在廣為人知的一點，稍後我會再多談一些。但我們的模式也致力於消除存取計算資源的任何障礙（barriers to access）。我們希望研究人員和創新者能在幾小時內存取 GPU，而不是幾週或幾個月。通常最長的延遲其實只是律師簽署協議的時間，但除此之外，存取啟動非常快，我們能在幾分鐘內開通。

Our value proposition is that we are a one-stop shop for advanced computing, especially advanced computing that is going to accelerate research where AI can make a positive impact on different research areas. So we offer cutting-edge compute. Uh, this is something that we are now well known for, and I’ll talk a little bit more about that. Uh, but also our model is to remove any barriers to access that compute. So we want researchers and innovators to be able to access GPUs in hours and not weeks and not months—the longest it usually takes is actually just for lawyers to agree to sign the agreements. But otherwise, it’s really quick. We can turn on access in minutes.

客戶選擇我們的另一個原因是安全性和合規性（security and compliance）。因為我們從一開始就知道，有些使用案例（use cases）涉及資料主權（data sovereignty），這一點非常重要，同時商業客戶也有很高的安全要求（security requirements）。我們主動制定了非常強大的安全和資料合規性路線圖（roadmap around security and data compliance），這是我們的另一個差異化優勢（differentiator），也是人們選擇我們的原因。當然，我們還有專業知識（expertise），包括硬體計畫（hardware program）中的 GPU 和基礎設施相關的一切，以及一些領域專長，例如如何在大規模叢集（large cluster）上進行藥物發現（drug discovery）或量子模擬（quantum simulations）。這些專業知識來自我們丹麥 AI 創新中心（DCAI）團隊，也來自我們的合作夥伴，包括 Nvidia 和其他我們能從社群中汲取的專家。

Another reason customers come to us is because of the security and compliance. So because we knew from the beginning that we’re going to have some use cases for data, sovereignty was very important, as well as commercial customers that have very high security requirements. We took it upon ourselves to have a very strong roadmap around security and data compliance. And that’s another differentiator for us and another reason people come to us. And of course, we have the expertise—and that’s the expertise in hardware programs, these GPUs, everything related to the infrastructure, but also some domain expertise in how to do drug discovery use cases on a large cluster or how to do quantum simulations on a large cluster. And that comes from us, the DCAI team, but also our partners, including Nvidia and other experts that we can tap into in the community.

第二個部分是實際建造這台電腦，這從夏天開始正式展開。我們建造的是一台大規模的 DGX 超級叢集（DGX SuperPOD）。希望在這場主題演講後，你們都能很熟悉 DGX 超級叢集是什麼。我們目前還是基於 H100 GPU，有 191 個 DGX 系統，每個系統配備 GPU，這就是我們得出最終數量的來源。

So the second piece was actually building the computer, and that started in earnest in the summer. Uh, it’s a large-scale DGX SuperPOD. Um, so hopefully after this keynote, you’re all very familiar with what the DGX SuperPOD is. Uh, we are still based on the H100 GPUs. Uh, we have 191 DGX systems. Uh, there’s GPUs for DGX systems, which is how we get that final number.

我們非常慎重地選擇了架構（architecture），因為這是 Nvidia 的參考架構（reference architecture）。我們知道我們希望這個投資能夠面向未來（future-proof），所以有了這個架構，我們可以輕鬆擴展（scale）和升級（upgrade），這也是我們的目標。目前，這台超級電腦（supercomputer）在全球 Top 500 排行榜中位列第 21 名。去年十月啟動後，我們做的第一件事就是在十一月進行基準測試（benchmarked）。我們很高興它在榜單上排名這麼高。這需要對 GPU 進行大量優化（optimization）才能達到這個成績，但我們對它的效能（performance）感到非常驕傲，每秒達 66.6 千萬億次浮點運算（66.6 Petaflops）。

We picked the architecture very deliberately because it’s a reference architecture for Nvidia. And we knew we wanted to future-proof the investment, and so with this, we can scale, we can upgrade fairly easily, and that’s the intent. So Gefion is the 21st fastest supercomputer in the Top 500 list. One of the first things that we did after the launch in October was to get it benchmarked in November. And we’re very pleased that it ranked so high on that list. Um, it took a lot of optimization of the GPUs to get there. But we’re very proud of that performance, which is 66.6 Petaflops per second.

我們還為它擁有全球第七快的儲存系統（storage system）感到驕傲，這一點也在 Top 500 榜單中有所體現。這要感謝我們的合作夥伴 DDN。我們為此感到驕傲，因為我們打造的是一個針對 AI 優化的系統（AI-optimized system）。對我們來說，這意味著我們不應該只專注於計算機本身，而是要提供一切能支持大型 AI 工作負載（large AI workloads）的東西，這是關鍵部分之一。

We’re also very proud of the fact that it is the 7th fastest storage system on the Top 500 list. Uh, and this is thanks to our partners at DDN. And we’re very proud of that because we are an AI-optimized system. And so for us, that meant that we shouldn’t focus on just the computer but offer everything that is going to enable large AI workloads, and that was a key part of it.

所以我們有來自 Nvidia 的頂級硬體（premium hardware），但我們也有 Nvidia 的專業軟體（specialized software）。這也是我們的客戶非常欣賞的一點，他們能獲得完整的軟體堆疊（software stack）以及一些專業框架（frameworks）和軟體，包括用於量子計算（quantum computing）的 cuQuantum，還有像 BioNeMo 這樣用於藥物發現（drug discovery）的工具。

So we have the premium hardware from Nvidia, um, but we also have specialized software from Nvidia. And that is also something that our customers have really been appreciating—having the entire software stack as well as some of the specialized frameworks and software, including cuQuantum for quantum, but also things like BioNeMo for drug discovery.

在今早的主題演講後，你們可能會覺得，嗯，相比 Jensen 講到的那些數字，我們的 Gefion 可能顯得微不足道（small potatoes）。但要記住我們在丹麥的目標。如果把所有可用的計算資源（compute）加起來，包括所有用於資助大規模計算的項目，我們總共大約有 4300 兆次浮點運算（4300 teraflops）。這已經是個不小的數字了。但相比之下，我們帶來的 Gefion 讓全國研究人員可用的總計算能力（total computing power）增加了 15 倍，這是一個巨大的提升（huge increase），也解釋了為什麼我們獲得了這麼多關注。

So after this morning’s keynote, you may think, well, maybe Gefion is small potatoes compared to some of the numbers that Jensen was throwing around. Uh, but one thing to keep in mind is what we’re trying to do here for Denmark. So if you take all of the available compute—so all of the initiatives that had gone into funding large-scale compute—we get to about 4,300 teraflops. And that is already a fairly big number. But as you can see, that looks fairly small compared to what we bring with Gefion, which was a 15X increase in the total computing power available in the entire country for researchers. Um, so a huge increase and also explains why we’ve been getting so much attention.

我想提到的另一個關於 AI 優化的部分是我們的 GPU 網路（networking）。這一點也值得一提。我們使用的是真正的 Infiniband（InfiniBand）來連接我們的 GPU。這意味著 GPU 之間的連線速度非常快（fast connectivity），這真的很重要。我們在與一些研究人員討論他們的工作負載（workloads）類型時發現，連線問題（connectivity issues）有時會讓處理過程變得非常緩慢。所以這也是我們優化的一部分。

Another thing I want to mention around AI optimization is the networking in our GPUs is something worth mentioning. Uh, we are using the actual InfiniBand for our GPUs. Um, and that means that we have very fast connectivity between GPUs, which is really important. That had also been a bottleneck when we talked to some researchers about the types of workloads that they were doing. That sometimes just took forever because of connectivity issues. So that is also something that we optimized.

於是，有趣的部分開始了——我們開始讓客戶上線（onboarding customers）。從一開始，我們就希望這個平台非常多功能（versatile）。我們不想讓這台電腦只服務於生命科學（life sciences）或量子研究（quantum research）。我們真的希望能支持多種類型的工作負載和多種類型的客戶。所以我們非常刻意地打造了一個與應用無關（application-agnostic）的電腦，這也意味著我們需要做更多工作，因為我們必須確保它能兼容各種各樣的使用案例（use cases）。但我們聽到用戶對某些領域特別感興趣，並在 Gefion 上運行這些應用。例如，自然語言處理（natural language processing）就很受關注，這可能並不令人意外。當人們聽到大規模超級電腦（large-scale supercomputer）時，會想到大模型訓練（training large models），這很合理。但我們也進行了一些很有趣的對話，關於在不同類型的資料（data）、影片（video）和音訊（audio）上進行分析和訓練。

So the fun part—we started onboarding customers. And from the beginning, we had this notion that we wanted the platform to be very versatile. We did not want to be a computer just for life sciences or a computer just for quantum research. Uh, we really wanted to enable multiple types of workloads and multiple types of customers. Um, so we very deliberately built the computer in a way that was application-agnostic, which also meant more work for us because we had to make sure that we were compatible with a wide variety of use cases. Uh, but these are some areas that we heard that users had a lot of interest in and running on Gefion. So we had a lot of interest in natural language processing, which is probably not that surprising. When people hear a large-scale supercomputer, they think LLMs and training large models, um, so that one made a lot of sense. Uh, but we also had some very interesting conversations around doing analysis and training on different types of data and video and audio.

這些領域經常被提及，因為在這些地方，要嘛模型（models）數量不多，要嘛現有模型的效果不夠好，存在一些有趣的研究空白（gap）。生物醫學研究（biomedical research）也經常被提到，從蛋白質結構預測（protein structure prediction）到生物醫學資料（biomedical data）的資料處理（data processing），都有涉及。模擬（simulation）也是一個越來越多人感興趣的領域，這其實超出了我最初的預期。當我們剛開始時，我們真的以為主要是訓練（training）會占大頭，但模擬也是一個很大的需求。

All that came up quite a bit as areas where there hadn’t been as many models or the models were not as good. And there was a gap in doing something interesting. Uh, biomedical research also comes up a lot—anything from protein structure prediction to doing data processing on biomedical data. And simulation is also something that a lot more people are interested in doing than I actually had assumed. Uh, originally, when we started, we really thought it was going to be a lot of training. Uh, the simulation is also a big need.

我們啟動時，真的很想盡快把這台電腦交到客戶手中，看看他們能用它做什麼。所以我們向社群發出了一個公開徵集（open call），請大家提出能在 Gefion 上運行的大膽想法（big ideas）。我們沒有指定他們應該專注於哪些領域，也沒有規定必須使用多少 GPU 或其他限制，我們真的希望他們能自由發揮創意（be creative）。但我們確實刻意挑選了不同類型的客戶。我們選了兩家新創公司（startups）——Cheetah 和 Go Autonomous，還選了一家公共部門機構 GMI，然後從不同大學選了一些學術研究者（academics）。

So when we launched, uh, we really wanted to put the computer in the hands of customers as soon as possible and see what they could do with it. So we made this open call to the community and asked them for big ideas about what interesting things they could run on Gefion. Uh, we did not specify any areas that they should focus on. Uh, we did not specify a certain number of GPUs they had to use or anything like that—we really wanted them to be creative. Uh, but what we did do was we deliberately picked different types of customers. So we picked two startups, Cheetah and Go Autonomous, and we picked GMI, which is a public sector organization. And then we picked some academics from different universities.

在研究方面，獲得一些非常有趣的項目真的很有意思，涵蓋了很多不同的領域。我們從丹麥技術大學（DTU）選了一個項目，研究新材料開發（new material development）和催化（catalysis）的新方法，這能幫助綠色轉型（green transition），以更永續的方式（sustainable way）進行化學反應。我們覺得這個項目很有趣，且有很高的影響潛力（high potential for impact），所以選了它。

So on the research side, uh, it was interesting to get some really interesting projects in a lot of different areas. Uh, and one of the ones we selected from DTU was around looking at new material development and new ways of doing catalysis that could help with the green transition, that could help chemical reactions in a more sustainable way. So we thought that was really interesting and had a high potential for impact. So we selected that.

我們還從哥本哈根大學（Copenhagen University）選了一個有趣的項目，研究多模態基因組資料（multimodal genome data）。他們結合了來自不同 DNA 和 RNA 的資料，創建一個新的基礎模型（foundation model）。這對研究來說也可能產生非常大的影響（impact）。我們看到學術機構提出了非常有創意且潛在影響巨大的項目，而他們過去從未有機會在沒有這種計算能力（compute）的情況下進行這樣的訓練。

Uh, but we also had from Copenhagen University an interesting project looking to do multimodal genome data. So combining data from different DNA, RNA, and so on, and creating a new foundation model. Uh, and that could also be really, really impactful in terms of research. So we are seeing very creative and potentially big-impact projects coming up from the academic institutions. And they had never been able to do this type of training without the type of compute that we bring to the table.

我們也在推動有影響力的創新（impactful innovation），不只是有影響力的研究（impactful research）。這一點在我們試營運階段（pilot phase）與兩家新創公司合作時得到了體現。例如，Go Autonomous 有一個宏大的願景，要消除重複性任務（repetitive tasks）。他們認為，任何重複的人類任務基本上都不應該由人來做，這些任務非常適合交給 AI，這樣人類就能專注於更有趣的事情。

So we’re also enabling impactful innovation in addition to impactful research, uh, and that was exemplified by the two startups that we worked with during this pilot phase. So Go Autonomous, for example, had this big vision of getting rid of repetitive tasks. So anything that is a repetitive human task basically should not be done by humans. It is well suited to be done by AI so that humans can focus on doing something more interesting.

他們決定研究的首要任務之一，是能夠識別資訊並在多個平台之間來回傳輸這些資訊（transfer information back and forth）。這是在自動化一個流程（automating a process），這個應用對於許多希望進行業務轉型（business transformation）和業務流程自動化（automation of business processes）的公司來說非常有吸引力。他們從去年十二月底開始使用這個平台，已經取得了相當不錯的成果（quite good results）。

Um, and one of the first things that they decided to look at is being able to identify information and transfer that information back and forth into multiple platforms. So automating a process that has a lot of applications also interesting to a lot of companies looking to do business transformation and automation of business processes. Uh, and they’ve been on the platform since late December and getting quite good results.

另一邊的光譜則與業務關係不大，而是更多關於醫療保健（healthcare）。Cheetah 也在做一些非常有趣的事情，他們的瓶頸在於想在大量的影片資料（video data）上進行訓練，這因為種種原因很具挑戰性。

On the other side of the spectrum, less about business and more into healthcare, uh, Cheetah is also doing something really interesting where their roadblock was that they wanted to train on a large amount of video data, uh, which is challenging to do for a lot of different reasons.

他們已經能在 Gefion 上做到這一點，他們這樣做的原因是試圖打造一個系統，能夠識別即將跌倒的病人（identify a patient about to have a fall）。這很難預測（predict），而且單純派一個人站在那裡看著病人很昂貴，也有些奇怪，對吧？想像有人只是等著病人跌倒然後去扶他們。所以這是一個在醫療保健（healthcare）場景中使用 AI 的很棒的方式，對病人有巨大的好處（huge benefit），只要做得好。他們從去年十二月底到今年一月初就開始與我們合作。

And they have been able to do that on Gefion. The reason that they are doing that is because they’re trying to have a system where they can identify a patient that’s about to have a fall. Uh, and that’s really hard to predict, and it can be quite costly to just have a human there watching a patient—also a little strange, right? Uh, just having somebody waiting for somebody to fall and to catch them. So this is a really nice way to use AI in that healthcare setting, which just has a huge benefit to the patients as long as it’s done well. And they’ve also been working with us since the end of December, early January.

我們在這些試營運項目（pilots）中取得了很好的成功。現在我們當然希望向更多客戶敞開大門，擴大 Gefion 的影響力（broaden the impact）。對我們來說，一個重要的目標是在丹麥的生命科學產業（life sciences industry）產生影響。這在丹麥已經是一個很大的產業，我之前提到的那些大公司都在這個領域。生命科學產業是丹麥經濟（Danish economy）的重要組成部分。

So we’ve had some really good successes with these pilots. And now of course we want to open the door to even more customers and broaden the impact that Gefion can do. One of the big things for us is going to be giving an impact on the life sciences industry in Denmark. And that is already a big industry in Denmark. Uh, we have all of these big companies that I mentioned earlier. And, um, the life sciences sector is a huge part of the Danish economy.

根據一些早期客戶的反饋，我們認為未來會看到的包括打造一個健康資料叢集（health data cluster）。我們可以利用 Gefion 是丹麥本土主權系統（sovereign system）的優勢，它高度安全（highly secure）。我們希望那些擁有非常敏感資料（sensitive data）和極高合規性要求（high compliance requirements）的客戶會來使用 Gefion，運行這些工作負載（workloads），並首次獲得過去無法得到的洞見（insights），因為之前他們沒有這樣的計算能力（compute）。這對我們來說是終極目標（holy grail）。

So what we think we’re going to see on Gefion, based on some early customers, includes, of course, this notion of having a health data cluster where we can use the fact that Gefion is a sovereign system in Denmark, highly secure. And we are hoping that customers with very sensitive data and very high compliance requirements will come to Gefion and run these workloads on us and be able to, for the first time, get insights that they couldn’t before because they couldn’t do that compute before. Um, so that is the holy grail for us.

但即使沒有健康敏感資料（health-sensitive data），我們也看到了非常有影響力的研究（impactful research）和使用案例（impactful use cases）。很多人對輔助藥物發現（assisted drug discovery）很感興趣，利用 AI 來協助藥物發現過程（drug discovery process），既提高效率（more efficient），也壓縮從分子（molecule）到可用於病人的藥物的時間線（timelines）。AI 可以在這個過程的每個階段發揮作用，我們看到人們對每個步驟都感興趣。

But we’re also seeing really impactful research and impactful use cases even without the health-sensitive data. There’s a lot of interest in assisted drug discovery. So using AI to assist in the drug discovery process, both to make things more efficient and also compress the timelines of being able to go from a molecule to a drug that can be used in a patient. So you can use AI really at all the different parts of that process. And we’re seeing interest in all the steps of that process.

計算醫學（computational medicine）也是一個非常有趣的領域，能夠基於基因組資料（genome data）和其他類型的資料開發新模型（new models），一路推進到個人化醫療（personalized medicine），根據遺傳特徵（genetic profiles）或某些疾病預測（predictions）改善治療方案（treatments）。這對我們來說是一個非常大的領域。

Uh, computational medicine is also something that’s really interesting—being able to develop new models based on genome data and other types of data and go all the way to personalized medicine and having improved treatments based on genetic profiles or predictions of certain diseases, uh, predictions. So, uh, a very big area for us.

另一個對我們來說非常重要的領域是量子技術（quantum）。我在這裡停一下，因為我一直在跟你們講一台 AI 超級電腦（AI supercomputer），你們可能會想，為什麼現在要講量子技術呢？

Another really big area for us is quantum. And here I’m going to pause because I’ve been talking to you about an AI supercomputer. So you may be thinking, why are you talking to us about quantum?

量子技術（quantum technology）和 AI 這兩個學科雖然不同，但它們在很多方面確實有重疊（overlap）。其中一個重疊的地方在於它們能影響的領域。量子技術和 AI 都對化學（chemistry）、物理學（physics）以及新演算法的開發（development of new algorithms）有很大的潛力。所以，像 Gefion 這樣的 AI 超級電腦（AI supercomputer）能做到的一件事，就是利用這台電腦來模擬量子系統（simulate quantum systems）。

Um, the two disciplines are different, but they do overlap in quite a few ways. Uh, and one of the ways that they overlap is in the types of things that they can impact. Uh, so both quantum and AI have a lot of potential for things like chemistry and physics and the development of new algorithms in those areas. So one of the things that could be done with an AI supercomputer, like Gefion, is actually using the computer to simulate quantum systems.

在量子電腦（quantum computers）的硬體成熟到足以開發量子演算法（quantum algorithms）之前，還有很多事情可以做。某種程度上，我們正在彌合這個差距（bridging the gap）。這真的很有趣。此外，AI 還有潛力幫助控制和優化（control and optimize）量子電腦的運行方式。這也是一個非常有趣的領域，我們一直在與量子計算公司（quantum computing companies）對話，探討他們如何利用 AI 更好地管理系統（manage their systems）或進行更好的量子錯誤校正（quantum error correction）。

Um, so there’s a lot that can be done before quantum computers—before the quantum hardware actually gets to the maturity level that is needed to develop a quantum algorithm. In some ways, we’re bridging the gap. Um, so that’s really interesting. And there’s also potential for AI to help control and optimize the way that quantum computers run. And that’s also a very interesting area where we’re constantly talking to quantum computing companies and figuring out how they could use AI to manage their systems better or do better quantum error correction, for example.

舉個非常具體的例子（specific example），我們一直在與哥本哈根大學（University of Copenhagen）的一個團隊合作，還有丹麥國家量子計算計畫（NQCP）的尼爾斯·玻爾研究所（Niels Bohr Institute）以及其他許多單位。他們研究的是如何打造一個系統，能夠非常精準地模擬量子（simulate quantum），以便測試量子演算法（quantum algorithms）。

So to give you a very specific example on this, um, we’ve been working with a team from the University of Copenhagen, uh, along with the NQCP Niels Bohr Institute and many others. Um, what they’ve been looking at is putting together a system that simulates quantum so well that they can test a quantum algorithm on it.

他們之前一直在用 CPU 做這個，甚至用了非常大規模的 CPU，但還是因為所需的計算量（amount of compute）而受限。當你在傳統電腦（classical computer）上模擬量子系統（quantum system）時，一個問題是資源需求呈指數級增長（increase exponentially）。所以如果你想模擬更多的量子位元（qubits），很快就達到了 CPU 的極限，無論你能買多少或擁有多少 CPU 都不夠。但用 AI 就不一樣，資源需求的增長曲線（curve）不會那麼陡峭。

And they had previously been doing this on CPUs, and even using very large-scale CPUs and still being blocked by the amount of compute that they needed. So one of the issues when you’re trying to simulate a quantum system on a classical computer here is the number of resources increases exponentially. So very quickly, if you’re trying to simulate more qubits, it very quickly gets to the limit of the number of CPUs you can reasonably buy or have at your disposal. With AI, that’s not the case—the curve doesn’t go up that high.

他們第一次使用 AI 系統時，用的是 Gefion。他們證明，與之前使用的大型 CPU 叢集（CPU cluster）相比，在 Gefion 上實現了 100 倍的加速（100 times speedup）。底下的資料是全新的，顯示他們現在能模擬比以前多得多的量子位元（qubits），這對他們的研究非常有影響力（impactful），因為他們試圖預測量子演算法的運作方式（how quantum algorithms work）。

So they’ve been using an AI system for the first time—they’ve been using Gefion—and what they were able to show was that they saw a 100 times speedup using Gefion over the CPU cluster, which was already very large, that they were using previously. The data at the bottom is actually brand new and it shows that they’re able to simulate a much higher number of qubits than before, which is very impactful for their research because what they’re trying to do is predict how quantum algorithms will work.

如果他們能模擬更高數量的量子位元（qubits），預測結果就會更接近實際情況，也就是當他們能在真正的量子電腦（quantum computer）上測試時會發生什麼。他們在過去幾週一直在做這件事，並持續調整這些結果（tweak these results），希望得到更好的表現。我得提一下，他們做到這一切並獲得如此巨大的進步時，甚至還沒用上整台電腦的全部能力。他們只用了我們大約三分之一的 GPU，因為我們對所有試營運客戶（pilot customers）都設了限制。但未來某個時候，我們可能會解除這個限制（remove that cap），讓他們用上整台電腦，看看他們能做到什麼。

So if they’re able to predict and simulate it on a higher number of qubits, then it more closely looks like what it will actually—what will happen when they are able to test it on a quantum computer. Um, so they’ve been doing this for the last few weeks and are continuing to tweak these results and get even better ones. One thing I should mention is they did all of this and had this huge improvement, and they’re not even using the full size of the computer. Uh, they’re using only about one-third of the number of GPUs that we have because we were throttling all of the pilot customers. So at some point, we may just remove that cap and let them take over the whole computer and see what they can do.

另一個非常有趣的使用案例（use case），關於當你從原本只有 CPU 轉而擁有大量 GPU 時能產生的影響，是來自丹麥生物研究所（Danish Biological Institute）的 GMI。他們的挑戰是試圖開發一個基於 AI 的天氣模型（AI-based weather model），看看它能否做出與傳統模型（traditional models）一樣好的預測。傳統模型是基於物理學（physics-based）的，雖然相當準確（fairly accurate），但需要大量計算能力（computing power）來運行，這意味著你不能經常運行它們。

Another really interesting use case in terms of the types of impact that can happen when you have a large number of GPUs when you only previously had CPUs is GMI, the Danish Biological Institute. And their challenge was trying to see if they could develop an AI-based weather model that could make predictions just as good as traditional models. Uh, traditional models are physics-based, and they are fairly accurate, but they take a lot of computing power to run, and that means you can’t run them as often.

GMI 有一個願景，希望利用機器學習（machine learning）和 AI 來實現他們的使命，也就是以某種方式預測天氣（predict the weather），確保社會受益（society benefits），並防止天氣對大量人口造成危險的大型事件（large-scale events）。一個很酷的事情是我很榮幸能宣布，GMI 在試營運階段的結果出來後，剛剛與我們簽訂了一份長期協議（long-term agreement）。

And GMI had this vision of being able to use machine learning and AI to accomplish their mission, which was to predict the weather in such a way that they could make sure that society benefits and that prevents large-scale events where weather can be very dangerous for a large segment of the population. So, uh, one really cool thing is I have the privilege of actually announcing the fact that GMI has just signed a long-term agreement with us following the results of this pilot.

基本上，在僅僅幾個月的試營運階段（pilot phase）中，他們已經發現他們開發的新模型幾乎與過去 20 年使用的傳統模型一樣準確。但現在這個模型效率高得多（much more efficient），意味著他們能更快、更頻繁地運行它（run it faster and a lot more often），從而得到更準確的預測（more accurate predictions）。他們證明了這一點適用於溫度和風速（temperature and wind）。他們對結果非常滿意，所以我們很高興能與他們在未來繼續合作。

Um, so basically, during the pilot phase, which only lasted a couple of months, they were already able to find that the new model that they developed was almost as accurate as the traditional model that they had been using for the last 20 years. Uh, but now the model is much more efficient, which means they can run it faster and run it a lot more often, which means more accurate predictions. They were able to show this for both temperature and wind. Uh, and they were very happy with the results. And so now we are very pleased that we get to work with them in the future.

他們將擴展這個模型，加入更多變數（variables），像是雲層覆蓋（cloud cover）之類的東西，並繼續訓練這個模型（keep training the model）讓它更準確。這真的是個好消息，也是我們第一個真正的大客戶（big customer），我們為此感到非常驕傲。有人問我，我們選這個作為第一個公布的客戶是不是因為天氣在丹麥是個大事。我很想說我們計畫得這麼完美，但其實不是。這只是很巧妙地湊在一起，他們有一個很棒的團隊一直在努力尋找用 AI 的方法，我們則提供了很棒的計算能力（compute）。

And they get to expand this model to include more variables like cloud cover and other things. Uh, and they get to keep training the model to make it even more accurate. Um, so that’s really a great piece of news and one of our first really big customers, and we really are proud about it. Uh, a couple of people have asked me if the reason we picked this as our first customer to announce is because the weather is such a big thing in Denmark. Uh, I’d like to say that we planned it that well, uh, but that is not the case. It’s just really worked out very nicely that they had this great team that had been working in trying to identify ways to use AI, and we were able to provide that wonderful compute.

我們在 GTC 這裡也有 DMI 的代表。如果你想了解更多關於這個模型以及他們做到的事情，應該在會議後留下來聽聽。所以我希望這能讓你們感受到客戶目前用這台在丹麥的超級叢集（superpod）做了什麼。我們的故事才剛開始，對吧？我給你們講了前四章（first four chapters），希望還有更多章節。我們會繼續努力，尋找不同的方式來加速未來（accelerate the future）。

Uh, and we actually have DMI here at GTC. So if you want to learn more about this model and what they’ve been able to do, you should stick around after the session. So I hope that gives you a sense of what customers have been able to do so far with a SuperPOD available in Denmark. We’re only at the beginning of the story, right? So I told you the first four chapters—hopefully there are many more. Uh, we continue to sell, and we’re continuing to find different ways that we can accelerate the future.

如果你有任何問題，或者感興趣的話，當然可以透過電子郵件聯繫我們（Email address）。我想我們還有時間做問答（Q&A），看看 Nick。好的，謝謝大家！

Um, so if you have any questions or if you’re interested, of course you’re free to contact us at the email address. And I think we have time for Q&A—looking at Nick. Yeah, thank you. Thank you, guys.

我們確實有時間回答一些問題。如果有人有問題，請舉手，我會把麥克風拿給你。這太棒了。我賭的是——你覺得怎麼樣？

And we do have time for some questions. So if anybody has a question, put your hand up, I’ll bring you the mic. Yeah, uh, this is awesome. I’m betting on—um, how are you finding?

你們還是第七快的儲存提供者（storage provider），有這麼多更快的儲存系統，對吧？

You are the 7th fastest storage provider as well, so many faster storage, yeah?

最快的。好的，嗯，你覺得 EU AI 法案（EU AI Act）未來會怎麼影響你們？

Fastest. Okay, um, how are you finding the EU AI Act going to be affecting you going forward?

我覺得它對我們的影響不大，反而可能更多影響我們的客戶。這是我們需要考慮的一個政策相關問題（policy-related issue）。一般來說，不確定性（uncertainty）會影響技術的採用速度（rate of adoption），尤其是對大企業來說，這可能決定他們是現在就大舉投資 AI，還是等三個月，直到他們的政策團隊說沒問題。我們確實看到一些非常大的公司有些猶豫（hesitation）。但總的來說，到目前為止，我們還沒看到太大影響，因為大家對運行這些大型工作負載（large workloads）有很強烈的意願和渴望，而過去的障礙一直是計算資源的可得性（availability of compute）。謝謝。

I don’t think it affects us so much as it may affect our customers. Um, and so that’s something to consider, uh, generally, with these policy-related issues. Uh, it’s the uncertainty that affects the rate of adoption, um, and especially with large enterprises, that could be the difference between them making a huge investment in AI now versus waiting 3 months until their policy people say, you know, it’s okay. Um, so we’ve seen a little bit of hesitation from the very large companies. But I would say for the most part, um, so far, we haven’t seen an impact because there was such a hands-up willingness and desire to be doing these large workloads, and the roadblock had been the availability of the compute. Thanks.

我們這邊還有另一個問題。

We have another question over here.

是的，我是埃德溫（Edwin）。我的問題是，今天早上在 GTC，Jensen 談到了加速發展（speed up）和成果產出（speed out）。我知道你們還在早期階段，但你怎麼看這個發展軌跡（trajectory）？或許你也能談談升級週期（upgrade cycle）對你們的影響。謝謝。

Yes, I’m Edwin. And my question is, um, inasmuch as at GTC this morning, Jensen was speaking about speed up and speed out, I realize it’s early days for you all, but how do you perceive that trajectory? And maybe you can speak about perhaps the upgrade cycle as well as it impacts you all. Thank you.

是的，GTC 的好處之一是你能看到所有即將推出的酷東西。但壞處也是你看到這些酷東西後，會覺得自己還不到一歲就已經過時了。但我覺得這對我們來說是件好事，因為我們有不同類型的用戶和非常多樣的工作負載（workloads）。我們的客戶會很滿意長期使用 H100，這是一款很棒的 GPU，對很多人來說，這是他們能想像到的最好的東西，甚至超乎他們的期待。對一些客戶來說，擁有最新的產品是有意義的。我們在一開始做的架構決策（architectural decisions）很好，因為我們有這種靈活性（ability）。所以我當然很想拿到新一代的產品（new generation），我想我們會在實際可行的情況下盡快做到這一點。對我們來說，這取決於客戶需求（customer demand）。如果有多個客戶說他們想要最新的設備（latest chip），我們就很容易做出決定並提供。我想我們很快就會這麼做。

Yeah, I mean one of the good things about GTC is you get to see all the cool stuff that’s coming out. And one of the bad things about GTC is, you get to see all the cool stuff that’s coming out, and you feel like you’re already old when you’re not even one year old. Um, but I think it’s a great thing, I think for us, with having multiple types of users and very different types of workloads, um, there are customers who will be happy using the H100 for a long time, and it’s a very good GPU, and for a lot of people, well, it is the best that they ever imagined that they could possibly get their hands on. Um, and for some customers, it makes sense to have the latest. So I think the nice thing about the architectural decisions we’ve made in the beginning is we have that ability. So I would definitely love to get my hands on the new generation. And I think we’ll do that as soon as it’s practical. And, uh, for us, it’s based on customer demand. Um, so if we have more than one customer saying I want the latest chip, then it’s easy for us to decide it and make it available. And I think we’ll definitely do that very soon.

有趣。謝謝。還有其他問題嗎？

Interesting. Thank you. Any more questions?

你好

Hello

我有幾個問題。你之前提到北約（NATO）在丹麥。這是因為市場需求嗎？還是與我們使用他們的計算資源（computing resources）有關的關係？

A couple of questions. So you mentioned before NATO is in Denmark. Is this because it’s in the market? Is this our relationship with using their computing resources?

北約，嗯，那可能是加速器（accelerator）的一部分。那是我們深度科技加速計畫（deep tech accelerator）的一部分，有一個由北約資助的黛安娜加速器（Diana Accelerator）。我想你指的是那張幻燈片吧。

Uh, NATO—uh, that could have been the accelerator. Uh, that’s part of our deep tech accelerator. There’s a Diana Accelerator that’s funded by NATO. Um, I think if that’s the slide I’m thinking you’re referring to.

是的，那是一個。謝謝。

Yeah, yeah, that’s one. Thank you.

你花了多少時間來部署整個系統（deploy the whole thing）？

How much time did you spend to deploy the whole thing?

實際建造系統（building the system）只花了幾個月，因為前幾個月有相當多的規劃（planning）。然後電腦本身組裝得相對很快，在十月推出前不久完成。所以一切都進行得非常快，超級快。我們不僅在加速技術，也在加速時間表（timeline acceleration）。但這需要所有合作夥伴齊心協力才能實現。Nvidia 在這個過程中表現得很出色。我們也有供應商交付了電腦。每個人都努力讓這個時間表實現。雖然這不容易，但值得去做。對了，回到之前的一個問題，這不是一個靜態的東西，對吧？你總是在升級（upgrading）或添加新東西。所以真正的答案是，建造這台電腦需要多久？你永遠不會真正完成。他們一直在調整和改進（tinkering）。

Uh, only a few months for the actual building of the system, uh, because there was quite a bit of planning in the first few months. And then, uh, the computer itself got assembled relatively quickly and shortly before launching in October. So everything happened really fast, uh, super fast. So we’re doing not just the acceleration, but timeline acceleration as well. Uh, but it took all of the partners coming together to make that happen. And Nvidia, of course, was amazing in that process. We had a vendor who also delivered the computer. And so everybody worked to make that timeline happen, but it is, um, yeah, it’s not easy, but it was worth making it happen. But there’s a, you know, to advance an earlier question. Um, it’s not a static thing, right? You’re always upgrading or you’re always adding things. So I think the real answer to how long does it take to build the computer? You’re never actually done. They’re always tinkering.

太棒了。還有其他問題嗎？

Awesome. And any more questions?

是的

Yes

你們在半年內就建造了它，這真的很令人印象深刻。我的問題是，它的效能是 6000 還是 66 千萬億次浮點運算（66 Petaflops）？它現在完全被占用了嗎？

And so you will build it in half a year. That’s really impressive. And the question is, uh, is it that it was 6,000 or 66 Petaflops? Is it fully occupied right now?

它的效能接近 66 千萬億次浮點運算（66 Petaflops），但還沒完全滿載，我們是故意這樣做的，因為我們還沒完全作為一個開放服務（open service）推出。我們保留了一些備用容量（spare capacity）。但根據目前的需求長度（length of demand），如果我們取消所有限制（remove all caps），很快就會滿載。

Uh, so it is close to, but not, uh, and we did that deliberately, uh, because we haven’t fully launched as an open service yet. Uh, we’re saving some spare capacity. Uh, but based on the length that we have, we would fill it up if we removed all the caps.

你有沒有考慮向更廣泛的歐洲客戶開放？

Do you think about opening up for the broader European customer?

我真的很高興你問這個問題，因為我應該早點提到，雖然我們的名字裡有「丹麥」（Danish），我們的使命之一是提升丹麥的生態系統（Danish ecosystem），但我們其實對丹麥以外的用戶是開放的。我們認為，對丹麥生態系統產生正面影響（positively impact）的最佳方式之一，就是允許來自歐洲和其他地方的客戶和用戶使用。當然，電腦的存取是完全遠端（completely remote）的，所以用戶可以從任何地方來。我們非常鼓勵歐洲的研究人員（researchers）和新創公司（startups）聯繫我們，爭取使用 Gefion 的時間。

I am really glad that you asked that because I should have mentioned that, uh, despite the fact that we have Danish in the name and of course part of our mission is to lift up the Danish ecosystem, we are actually open to users outside of Denmark. And we think some of the best ways to positively impact the Danish ecosystem is to allow customers and users from Europe and elsewhere, um, and of course the access to the computer is completely remote, so users can come from everywhere, and we definitely encourage European researchers and startups to reach out and get time on Gefion.

我想如果還有問題，我們還有時間再問一個。是的，很多問題都出於好奇。你有沒有收到過要求，比如用高品質資料（high-quality data）訓練一個丹麥語（Danish language）的模型？因為我們在其他國家看到這樣的趨勢，對吧？

I think we’ve got time for one more if there’s another question. Yes. Yeah, many questions out of curiosity? Have you got any requests to have, like, a model trained in Danish language with high-quality data? Because we have seen these trends in other countries, right?

大概一天會被問兩次吧，有人會來找我談這個。他們確實分享了一些想法（shared some ideas）。其實有很正當的理由去做這個。我不是丹麥人，自從八月接下這份工作後就住在丹麥，我可以保證，丹麥語是一種很難學又很特別的語言（tough and unusual language）。在很多領域，訓練一個專為丹麥語或至少日耳曼北歐語言（Germanic Nordic languages）的模型是有意義的。有幾個團隊正在思考怎麼最好地做到這一點，當然，只要他們準備好了，我們會支持他們。我們現有的基礎設施（infrastructure）非常適合這種工作。

Only about twice a day do I get approached about it? They did share some ideas. There’s actually some valid reasons to do it. Uh, I mean, I’m obviously not Danish, and I’ve been living in Denmark since August since I took this job, and I can vouch for the fact that Danish is a very tough language and a very unusual language. Uh, and there are a lot of areas where it makes sense to have something that’s actually trained on the Danish language or at least the Germanic Nordic languages. Um, there are a couple of groups that are thinking about how to best do that, and of course we can and will support them. As soon as they are ready, the type of infrastructure that we have is very well suited to do that type of work.

好的，太棒了，請大家再次熱烈歡迎今天跟我們分享的講者。謝謝！