(万字长文)- 黄仁勋昨晚在高盛科技大会上说了什么

文摘   2024-09-12 05:45   中国香港  


- 这两天高盛在 San Francisco 开 Communacopia + Technology Conference,各大科技公司大佬云集;


- 昨天的大会由英伟达老黄开场,万众瞩目;


- 会议是凌晨开的,直播还没看完,英伟达股价已经起飞了


- 直播一结束,彭博上面就已经有了英文的全文纪要(虽然有些地方颇为模糊)。


- 因为刚好听了录音,也挺有意思的,顺手整理一下几个东西,凑凑热闹。按顺序排列:-




1. 高盛做的大会 key takeways (省时间的朋友看)


2. 中文的一些梳理版原文(需要语境的朋友看)


3. 中英对照版原文全文(有兴趣研究细节的朋友,也可以翻翻)



录音就不方便直接放出来了;有兴趣的朋友再另外联系。


1. Key takeaways




2. 中文梳理版


1. 首先談談31年前,你創立公司時的一些想法。從那時起,你將公司從一個專注於遊戲的GPU公司轉型爲一個爲數據中心行業提供廣泛硬件和軟件的公司。你能不能先談談這個歷程?當你開始時,你在想什麼?它是如何演變的?你未來的關鍵優先事項是什麼,以及你如何看待未來的世界?


黃仁勳:我想說,我們做對的一件事是,我們預見到,未來會有另一種計算形式,它可以增強通用計算,解決通用工具永遠無法解決的問題。這種處理器一開始會做一些對CPU來說極其困難的事情,那就是計算機圖形處理。

但我們將逐步擴展到其他領域。我們選擇的第一個領域當然是圖像處理,這與計算機圖形處理是互補的。我們將其擴展到物理模擬,因爲在我們選擇的視頻遊戲領域中,你不僅希望它美觀,還希望它動態化,能夠創建虛擬世界。我們一步一步地擴展,並將其引入科學計算。第一個應用之一是分子動力學模擬,另一個是地震處理,這基本上是逆物理。地震處理與CT重建非常相似,是另一種形式的逆物理。所以我們一步一步地解決問題,擴展到相鄰行業,最終解決了這些問題。


我們一直堅守的核心理念是加速計算能夠解決有趣的問題。我們的架構保持一致,意味着今天開發的軟件可以在你留下的大量已安裝基礎上運行,過去開發的軟件可以通過新技術加速。這種關於架構兼容性的思維方式、創建大量已安裝基礎、與生態系統共同發展的心理從1993年就開始了,我們一直延續到今天。這就是爲什麼英偉達的CUDA擁有如此龐大的已安裝基礎的原因,因爲我們一直在保護它。保護軟件開發者的投資是我們公司自始至終的首要任務。

保護軟件開發者的投資是我們公司自始至終的首要任務。展望未來,我們在一路上解決的一些問題,當然包括學習如何成爲創始人、如何成爲首席執行官、如何經營業務、如何建立公司,這些都是新的技能。這有點像發明現代計算機遊戲行業。人們可能不知道,但英偉達是世界上最大的視頻遊戲架構的安裝基礎。GeForce擁有大約3億玩家,仍然在快速增長,非常活躍。所以我認爲,每次我們進入一個新市場時,我們都需要學習新的算法、市場動態,創建新的生態系統。


我們需要這樣做的原因是,與通用計算機不同,通用計算機一旦構建好處理器,所有的東西最終都會運行。但我們是加速計算機,這意味着你需要問自己,你要加速什麼?不存在所謂的通用加速器。


2. 深入談談一般用途和加速計算之間的差異?


黃仁勳:如果你看看現在的軟件,你寫的軟件中有大量的文件輸入輸出,有設置數據結構的部分,還有一些魔法般的算法核心。這些算法不同,取決於它們是用於計算機圖形處理、圖像處理還是其他什麼。它可以是流體、粒子、逆物理或者圖像領域的東西。所以這些不同的算法都是不同的。如果你創建一個處理器,專門擅長這些算法,並補充CPU處理它擅長的任務,那麼理論上,你可以極大地加速應用程序的運行。原因是通常5%到10%的代碼佔據了99.99%的運行時間。


因此,如果你把那5%的代碼卸載到我們的加速器上,技術上,你可以將應用程序的速度提高100倍。這並不罕見。我們經常可以將圖像處理加速500倍。現在我們做的是數據處理。數據處理是我最喜歡的應用之一,因爲幾乎所有與機器學習相關的內容都在演進。它可以是SQL數據處理、Spark類型的數據處理,或者是向量數據庫類型的處理,處理無結構或結構化的數據,這些數據都是數據幀。


我們對這些進行極大的加速,但爲了做到這一點,你需要創建一個頂級的庫。在計算機圖形處理領域,我們很幸運有了Silicon Graphics的OpenGL和Microsoft的DirectX,但在這些之外,沒有真正存在的庫。因此,舉個例子,我們最著名的一個庫是與SQL類似的庫。SQL是存儲計算的庫,我們創建了一個庫,它是世界上第一個神經網絡計算庫。


我們有cuDNN(用於神經網絡計算的庫),還有cuOpt(用於組合優化的庫),cuQuantum(用於量子模擬和仿真的庫),以及很多其他的庫,比如用於數據幀處理的cuDF,類似於SQL的功能。因此,所有這些不同的庫都需要被髮明出來,它們可以把應用程序中的算法重新整理,使我們的加速器能夠運行。如果你使用這些庫,你就可以實現100倍的加速,獲得更多的速度,非常驚人。


因此,概念很簡單,而且非常有意義,但問題是,你如何去發明這些算法,並讓視頻遊戲行業使用它們,編寫這些算法,讓整個地震處理和能源行業使用它們,編寫新的算法並讓整個AI行業使用它們。你明白我的意思嗎?因此,所有這些庫,每一個庫,首先我們必須完成計算機科學的研究,其次,我們必須經歷生態系統的開發過程。


我們必須去說服每個人使用這些庫,然後還要考慮它們運行在哪些類型的計算機上,每種計算機都不一樣。因此,我們一步一步地進入一個領域又一個領域。我們爲自動駕駛汽車創建了一個非常豐富的庫,爲機器人開發了一個非常出色的庫,還有一個令人難以置信的庫,用於虛擬篩選,無論是基於物理的虛擬篩選還是基於神經網絡的虛擬篩選,還有一個令人驚歎的庫用於氣候技術。


因此,我們必須去結交朋友,創建市場。事實證明,英偉達真正擅長的事情是創建新的市場。我們現在已經做了這麼久,以至於英偉達的加速計算似乎無處不在,但我們確實必須一步步地完成,一次一個行業地開發市場。


3. 現場的許多投資者非常關注數據中心市場,能否分享一下你對中長期機會的看法?顯然,你的行業推動了你所稱的「下一次工業革命」。你如何看待數據中心市場的現狀以及未來的挑戰?


黃仁勳:有兩件事同時在發生,它們經常被混爲一談,分開討論有助於理解。首先,我們假設沒有AI存在的情況下。在沒有AI的世界裏,通用計算已經停滯不前了。大家都知道,半導體物理學中的一些原理,比如摩爾定律、Denard縮放等,已經結束了。我們不再看到CPU的性能每年翻倍的現象。我們已經很幸運了,能在十年內看到性能翻倍。摩爾定律曾經意味着五年內性能提升十倍,十年內提升一百倍。


然而現在這些已經結束了,所以我們必須加速一切能加速的東西。如果你在做SQL處理,加速它;如果你在進行任何數據處理,加速它;如果你在創建一個互聯網公司並擁有推薦系統,必須加速它。如今最大的推薦系統引擎已經全部加速了。幾年前這些還在CPU上運行,而現在已經全部加速了。因此,第一個動態是,全世界價值數萬億美元的通用數據中心將會現代化,轉變爲加速計算的數據中心。這是不可避免的。


此外,因爲英偉達的加速計算帶來了如此巨大的成本降低,過去十年中,計算能力不是以100倍,而是以100萬倍的速度增長。那麼問題來了,如果你的飛機能快一百萬倍,你會做什麼不同的事情呢?


因此,人們突然意識到:「爲什麼我們不讓計算機來編寫軟件,而不是我們自己去想象這些功能,或者我們自己去設計算法呢?」我們只需要把所有的數據、所有的預測性數據交給計算機,讓它去找出算法——這就是機器學習,生成式AI。因此,我們在許多不同的數據領域大規模應用了它,計算機不僅知道如何處理數據,還理解數據的含義。因爲它同時理解多種數據模式,它可以進行數據翻譯。


因此,我們可以從英文轉換爲圖像,從圖像轉換爲英文,從英文轉換爲蛋白質,從蛋白質轉換爲化學物質。因爲它理解了所有的數據,因此可以進行所有這些翻譯過程,我們稱之爲生成式AI。它可以將大量的文字轉換爲少量的文字,或者將少量的文字擴展爲大量的文字,等等。我們現在正處於這個計算機革命的時代。


而現在令人驚歎的是,第一批價值數萬億美元的數據中心將被加速,並且我們還發明瞭這種新型的軟件,稱爲生成式AI。生成式AI不僅僅是一種工具,它是一種技能。正是因爲這個原因,新的行業正在被創造出來。


這是爲什麼?如果你看看直到現在的整個IT行業,我們一直在製造人們使用的工具和儀器。而第一次,我們正在創造出能夠增強人類能力的技能。因此,人們認爲AI將超越價值數萬億美元的數據中心和IT行業,進入技能的世界。


那麼,什麼是技能呢?比如數字貨幣是一種技能,自動駕駛汽車是一種技能,數字化的裝配線工人,機器人,數字化的客戶服務,聊天機器人,數字化的員工爲英偉達規劃供應鏈。這可以是一個SAP的數字代理。我們公司大量使用ServiceNow,我們現在擁有了數字員工服務。因此,我們現在擁有了這些數字化的人類,這就是我們現在正處的AI浪潮。


4. 金融市場中存在一個持續的辯論,即隨着我們繼續建設AI基礎設施,投資回報是否足夠?你如何評估客戶在這個週期中獲得的投資回報率?如果你回顧歷史,回顧PC和雲計算,它們在類似的採用週期中,回報率如何?與現在相比有什麼不同?


黃仁勳:這是個非常好的問題。讓我們來看看。在雲計算之前,最大的趨勢是虛擬化,如果大家還記得的話。虛擬化基本上意味着我們將數據中心中的所有硬件虛擬化爲虛擬數據中心,然後我們可以跨數據中心移動工作負載,而不必直接與特定的計算機相關聯。結果是,數據中心的利用率提高了,我們看到了數據中心成本減少了兩倍到兩倍半,幾乎是在一夜之間完成的。


接着,我們將這些虛擬計算機放到雲中,結果是,不僅僅是一家公司,很多公司都可以共享相同的資源,成本再次下降,利用率再次提高。


這些年的所有進步,掩蓋了底層的根本變化,那就是摩爾定律的終結。我們從利用率提升中獲得了兩倍、甚至更多的成本降低,然而這也碰到了晶體管和CPU性能的極限。


接着,所有的這些利用率的提升已經達到極限,這也是爲什麼我們現在看到數據中心和計算通脹的原因。因此,第一件正在發生的事情就是加速計算。因此,當你在處理數據時,比如使用Spark——這是當今世界上使用最廣泛的數據處理引擎之一——如果你使用Spark並通過英偉達加速器加速它,你可以看到20倍的加速。這意味着你會節省10倍的成本。


當然,你的計算成本會上升一點,因爲你需要支付英偉達GPU的費用,計算成本可能會增加一倍,但你將減少計算時間20倍。因此,你最終節省了10倍的成本。而這樣的投資回報率對於加速計算來說並不罕見。因此,我建議你們加速一切可以加速的工作,然後使用GPU進行加速,這樣可以立即獲得投資回報。


除此之外,生成式AI的討論是當前AI的第一波浪潮,基礎設施玩家(比如我們自己和所有云服務提供商)將基礎設施放在雲上,供開發人員使用這些機器來訓練模型、微調模型、爲模型提供保護等等。由於需求如此之大,每花費1美元在我們這裏,雲服務提供商可以獲得5美元的租金回報,這種情況正在全球範圍內發生,一切都供不應求。因此,對這種需求的需求非常巨大。


我們已經看到的一些應用,當然包括一些知名的應用,比如OpenAI的ChatGPT、GitHub的Copilot,或者我們公司內部使用的共同生成器,生產力提升是不可思議的。我們公司裏的每一個軟件工程師現在都使用共同生成器,不管是我們自己爲CUDA創建的生成器,還是用於USD(我們公司使用的另一種語言),或者Verilog、C和C++的生成器。


因此,我認爲每一行代碼都由軟件工程師編寫的日子已經徹底結束了。未來,每一個軟件工程師都將有一個數字工程師伴隨在身邊,24/7隨時協助工作。這就是未來。因此,我看英偉達,我們有32000名員工,但這些員工周圍將有更多的數字工程師,可能會多100倍的數字工程師。


5. 很多行業都在接受這些變化。哪些用例、行業是你最興奮的?


黃仁勳:在我們公司,我們在計算機圖形學方面使用AI。如果沒有人工智能,我們無法再進行計算機圖形學。我們只計算一個像素,然後推測其餘的32個像素。也就是說,我們在某種程度上「幻想」出其餘的32個像素,它們在視覺上是穩定的,看起來是照片級真實的,圖像質量和性能都非常出色。


計算一個像素需要大量的能量,而推測其他32個像素的能量需求則非常少,而且可以非常快速地完成。因此,AI並不僅僅是訓練模型,這只是第一步。更重要的是如何使用模型。當你使用模型時,你會節省大量的能量和時間。


如果沒有AI,我們無法爲自動駕駛汽車行業提供服務。如果沒有AI,我們在機器人技術和數字生物學領域的工作也是不可能的。現在幾乎每一個科技生物公司都以英偉達爲中心,他們正在使用我們的數據處理工具來生成新蛋白質,小分子生成、虛擬篩選等領域也將因爲人工智能而被徹底重塑。


6. 談談競爭和你們的競爭壁壘吧。目前有很多公私公司希望能打破你們的領導地位。你如何看待你們的競爭壁壘?


英偉達:首先,我認爲有幾件事讓我們與衆不同。第一點要記住,AI並不僅僅是關於芯片的。AI是關於整個基礎設施的。如今的計算機不是製造一塊芯片然後人們購買它並放入計算機中。那種模式屬於上世紀90年代。如今的計算機是以超級計算集群、基礎設施或超級計算機爲名開發的,這不是一塊芯片,也不完全是計算機。


所以,我們實際上是在構建整個數據中心。如果你去看一下我們其中一個超級計算集群,你會發現管理這個系統所需的軟件是非常複雜的。並沒有一個「Microsoft Windows」可以直接用於這些系統。這種定製化的軟件是我們爲這些超級集群所開發的,所以設計芯片的公司、構建超級計算機的公司以及開發這些複雜軟件的公司,理所當然的是同一家公司,這樣可以確保優化、性能和效率。


其次,AI本質上是一種算法。我們非常擅長理解算法的運作機制,並且了解計算堆棧如何分佈計算,以及如何在數百萬個處理器上運行數天,保持計算機的穩定性、能源效率以及快速完成任務的能力。我們在這方面非常擅長。


最後,AI計算的關鍵是安裝基礎(installed base)。擁有跨所有云計算平台和內部部署(on-premise)的統一架構非常重要。無論你是在雲中構建超級計算集群,還是在某台設備上運行AI模型,都應該有相同的架構以運行所有相同的軟件。這就是所謂的安裝基礎。而這種自1993年以來的架構一致性是我們能夠取得今天成就的關鍵原因之一。


因此,今天如果你要創辦一家AI公司,最明顯的選擇就是使用英偉達的架構,因爲我們已經遍佈所有的雲平台,不論你選擇哪臺設備,只要它有英偉達的標識,你就可以直接運行相同的軟件。


7. Blackwell在訓練上快了4倍,推理速度比它的前代產品Hopper快了30倍。你們的創新速度如此之快,你們能否保持這樣的節奏?你們的合作伙伴能否跟上你們的創新步伐?


黃仁勳:我們的基本創新方法是確保我們不斷推動架構創新。每個芯片的創新週期大約是兩年,在最好的情況下是兩年。我們每年還會對它們進行中期升級,但整體架構的革新大約是每兩年一次,這已經非常快了。


我們有七個不同的芯片,這些芯片共同作用於整個系統。我們可以每年推出新的AI超級計算集群,並且比上一代更強大。這是因爲我們擁有多個可以進行優化的部分。因此我們可以非常快速地交付更高的性能,並且這些性能的提升直接轉化爲總擁有成本(TCO)的下降。


Blackwell在性能上的提升意味着,對於擁有1千兆瓦電力的客戶,他們可以獲得3倍的收入。性能直接轉化爲吞吐量,吞吐量則轉化爲收入。如果你有1千兆瓦的電力可用,你可以獲得3倍的收入。


因此,這種性能提升的回報是無與倫比的,也無法通過芯片成本的降低來彌補這3倍的收入差距。


8. 如何看待對亞洲供應鏈的依賴?


黃仁勳:亞洲的供應鏈非常複雜並且高度互聯。英偉達的GPU不僅僅是一塊芯片,它是由成千上萬個組件組成的複雜系統,類似於一輛電動車的構造。因此,亞洲的供應鏈網絡非常廣泛且複雜。我們力求在每一個環節上設計出多樣性和冗餘性,確保即使出現問題,我們也能夠迅速將生產轉移到其他地方進行製造。總的來說,即使供應鏈出現中斷,我們也有能力進行調整,以確保供應的連續性。


我們目前在臺積電進行製造,因爲它是世界上最好的,不僅僅是好一點點,而是好得多。我們與他們有着長期的合作歷史,他們的靈活性和規模能力都令人印象深刻。


去年,我們的收入出現了大幅增長,這離不開供應鏈的快速反應。台積電的敏捷性以及它們滿足我們需求的能力是非常了不起的。在不到一年的時間裏,我們大幅提升了產能,並且我們明年將繼續擴大,後年還要進一步擴大。因此,他們的敏捷性和能力都很出色。不過,如果有需要,我們當然也可以轉向其他供應商。


9. 貴公司處於非常有利的市場位置。我們已經討論了很多非常好的話題。你最擔心的是什麼?


黃仁勳:我們的公司目前與全球每一家AI公司都有合作,也與每一家數據中心有合作。我不知道有哪家雲服務提供商或計算機制造商我們沒有合作的。因此,隨着這樣的規模擴展,我們肩負着巨大的責任。我們的客戶非常情緒化,因爲我們的產品直接影響他們的收入和競爭力。需求太大,滿足這些需求的壓力也很大。


我們目前正全面生產Blackwell,並計劃在第四季度開始發貨並進一步擴展。需求如此之大,每個人都希望能夠儘早拿到產品,獲取最多的份額。這種緊張和激烈的氛圍實在是前所未有。

雖然在創造下一代計算機技術時非常令人興奮,也令人驚歎地看到各種應用的創新,但我們肩負着巨大的責任,感到壓力很大。但我們盡力去做好工作。我們已經適應了這種強度,並將繼續努力。


3. 中英对照全文


采访人:


Good morning. Good morning.

早上好。早上好。


老黄:


Thank you. Good morning. Great to see everybody.

谢谢。早上好。很高兴见到大家。


采访人:


I flew in late last night. I didn't really expect to be on stage at seven, twenty in the morning, but you know, seems everybody else did. So here we are, Jensen. Thank you for being here. I'm delighted to be here. Thank you all for being here. I hope everybody's been enjoying the conference.It's, it's a fantastic event. Lots of great, lots of great companies, a couple thousand people here. And so really, really terrific and obviously a real highlight and a real privilege to have Jensen President, CEO video here.Since you found Nvidia 1993, you've pioneered accelerating computing. The company's invention of the gpu in 1999 sparked the growth.Of the PC gaming market, redefining computers and igniting the era of modern AI.Jensen holds a bse degree from Oregon state university and mse degree from Stanford, and so I want to start by welcoming you Jensen. Everybody, please welcome Jensen to the state to do this reallyually and I'm going to try to get you talking about some things and I know you're passionate about, but I just want to start.31 years ago founded the company.You've transformed yourself from a gamingcentric gpu company to one that offers a broad range of hardware software to the data center industry. And I'd just like you to start by talking a little bit about the journey when you started. What were you thinking, how it evolved? Because it's been a pretty extraordinary journey, and then maybe.You can break from that, just talk a little bit as you position forward on your key priorities and how you're looking at the world going forward.

我昨晚很晚才到。我没想到早上七点二十会在台上,但你知道,似乎其他人都这样认为。所以我们在这里,詹森。感谢你在这里。我很高兴能在这里。感谢大家的到来。希望大家都在享受这次会议。这是一个精彩的活动。很多优秀的公司,这里有几千人。所以真的非常棒,显然这是一个真正的亮点,能够有詹森总裁兼首席执行官在这里真是个特权。自 1993 年创立英伟达以来,你一直在推动计算加速。公司在 1999 年发明的 GPU 引发了 PC 游戏市场的增长,重新定义了计算机,并点燃了现代人工智能的时代。詹森拥有俄勒冈州立大学的学士学位和斯坦福大学的硕士学位,所以我想先欢迎你,詹森。大家,请欢迎詹森上台,我会尽量让你谈谈一些你热衷的事情,但我想先说,31 年前你创立了这家公司。您已经将自己从一家以游戏为中心的 GPU 公司转变为一家向数据中心行业提供广泛硬件和软件的公司。我希望您能先谈谈您开始时的旅程。您当时在想什么,这个过程是如何演变的?因为这是一段相当非凡的旅程,然后也许您可以从中抽离,谈谈您在未来的关键优先事项以及您如何看待未来的世界。


老黄:


Yes, David, it's good to be.The thing that we got right, I would say is, is that our vision, that there would be another form of computing that could augment.General purpose computing to solve problems that a general purpose instrument won't ever be good at.And that processor would start out doing something that was insanely hard for cpus to do. And it was computer graphics, but that we would expand that over time to do other things. The first thing that we chose, of course, was image processing, which was just complementary to computer graphics. We extended it to physics simulation because.In the domain, the application domain that we selected, the video games, you want it to be beautiful, but you also want it to be dynamic, to create virtual worlds. We took stepbystepbystep and we took it into scientific computing. Beyond that, one of the first applications was molecular or dynamic simulation. Another was seismic processing, which is basically inverse physics. Seismic processing is very similar to CT reconstruction, another form of inverse physics. And so we just took it step by step by step, reasoned about complementary types of algorithms. Adjacent industries kind of solved our way here, if you will.But the common vision at the time was that accelerated computing would be able to solve problems that are interesting.And that if we were able to keep the architecture consistent, meaning, have an architecture where software that you developed today could run on a large installed base that you've left behind. And the software that you created in the past would be accelerated even further by new technology. This way of thinking about architecture, compatibility, creating large installed base, taking the software investment of the ecosystem along with us. That psychology started in 1993 and we carried it to this day.Which is the reason why nvidiauda has such a massive installed base, and that because we always protected it, protecting the investment of software developers has been the number one priority of our company since the very beginning and.Um.Going forward.Some of the things that we solved along the way of course learning how to be learning how to be a founder, learning how to be a CEO, learning how to conduct a business, learning how to build a company, not to stuff these are, these are all, you know, new skills and and which is which is kind of like learning how to, how to invent the modern computer gaming industry. Nvidia people don't know this, but Nvidia is the largest installed base of video game architecture in the world. G force is some 300 million gamers in the world is still growing incredibly well, super vibrant.And so I think the every single time we had to go and enter into a new market, we had to learn new algorithms, new market dynamics, create new ecosystems. And the reason why we have to do that is because unlike a generalpurpose computer, if you built that processor.Then everything eventually just kind of works, but we're an accelerated computer, which means the question you have to ask yourself is what do you accelerate?There's no such thing as a universal accelerator because.

是的,大卫,确实很好。我们做对的事情,我会说是我们的愿景,即会有另一种计算形式,可以增强通用计算,以解决通用工具永远无法擅长的问题。那个处理器最初会做一些对 CPU 来说极其困难的事情,那就是计算机图形,但我们会随着时间的推移扩展到其他领域。我们选择的第一件事当然是图像处理,这与计算机图形是互补的。我们将其扩展到物理模拟,因为在我们选择的应用领域——视频游戏中,你希望它既美观又动态,以创造虚拟世界。我们一步一步地推进,最终进入了科学计算。除此之外,第一个应用之一是分子或动态模拟。另一个是地震处理,基本上是逆物理。地震处理与 CT 重建非常相似,都是另一种逆物理。因此,我们就是一步一步地推进,推理出互补类型的算法。相邻行业在某种程度上解决了我们来到这里的方式。但当时的共同愿景是,加速计算能够解决有趣的问题。如果我们能够保持架构的一致性,也就是说,拥有一个架构,使得您今天开发的软件能够在您留下的大量已安装基础上运行。而您过去创建的软件将通过新技术得到进一步加速。这种关于架构、兼容性、创建大型已安装基础、将生态系统的软件投资带入我们的思维方式始于 1993 年,并一直延续至今。这就是为什么 nvidiauda 拥有如此庞大的已安装基础的原因,因为我们始终保护它,保护软件开发者的投资自一开始就是我们公司的首要任务。嗯。展望未来。在这个过程中我们解决的一些问题当然包括学习如何成为创始人,学习如何成为首席执行官,学习如何开展业务,学习如何建立公司,这些都是新的技能,这有点像学习如何发明现代计算机游戏产业。人们不知道,Nvidia 是全球最大的视频游戏架构安装基础。G Force 在全球有大约 3 亿玩家,并且仍在快速增长,极其活跃。因此,我认为每次我们进入一个新市场时,我们都必须学习新的算法、新的市场动态,创造新的生态系统。我们之所以必须这样做,是因为与通用计算机不同,如果你构建了那个处理器,那么一切最终都会正常工作,但我们是加速计算机,这意味着你必须问自己,你加速的是什么。没有所谓的通用加速器。


采访人:


Yes, dig down on this a little bit different. Just talk about the differences between general purpose and accelerating.

是的,深入探讨一下这个稍有不同的内容。只是谈谈通用和加速之间的区别。


老黄:


Computing, if you look at, if you look at software.Um.Out of your body of software that you wrote, there's a lot of file io, there's setting up the data structure. There's, there's a part of the software inside which has.Some of the magic kernels, the magic algorithms, and these algorithms are different depending on whether it's computer graphics or image processing or whatever it happens to be. It could be fluids, it could be particles, it could be inverse physics. As I mentioned, it could be image domain type stuff. And so all these different algorithms are different and.If you created a processor, that is somehow really, really good at those algorithms.And you complement the cpu, where the cpu does whatever it's good at then.Theoretically, you could take an application and speed it up tremendously, and the reason for that is because.Usually, some 5%10% of the code represents 99~999% of the runtime.And so if you take that 5% of the code and you offloaded it on our accelerator, then technically you should be able to speed up the application 100 times. And it's not abnormal that we do that. It's not unusual.And so we'll speed up image processing by 500 times. And now we do data processing. Data processing is one of my favorite applications because almost everything related to machine learning, which is a datadriven way of doing software. Data processing is evolved. It could be sql data processing, could be spark type of data processing. It could be vector database type of processing, all kinds of different ways of processing, either unstructured data or structured data, which is data frames.And we accelerate the living daylights out of that. But in order to do that, you have to create that library, that dyn library on top. And in the case of computer graphics, we were fortunate to have silicon graphics as opengl and Microsoft directx.But outside of those, no libraries really existed. And so for example, one of our most famous libraries is a library, kind of like sql is a library, sql is a library for in storage computing. We created a library called kudnn. Kudnn is the world's first neural network computing library. And so we have qdnn we have.COO, opt for combinatory optimization. We have ku quantum for quantum simulation and emulation. All kinds of different libraries cdf for data, data frame processing, for example sql. And so all these different libraries have to be invented. That takes the algorithms that run in the application and refactor those algorithms in a way that our accelerators can run.And if you use those libraries, then you get 100 X.

计算,如果你看看,如果你看看软件。嗯。在你编写的软件中,有很多文件输入输出,有设置数据结构的部分。软件内部有一些魔法内核,魔法算法,这些算法根据是计算机图形学、图像处理或其他任何情况而有所不同。它可以是流体,可以是粒子,可以是逆物理。如我所提到的,它可以是图像域类型的东西。因此,所有这些不同的算法都是不同的。如果你创建了一个处理器,它在这些算法上非常非常出色。并且你补充了 CPU,CPU 做它擅长的事情。那么,从理论上讲,你可以将一个应用程序的速度大幅提升,原因在于。通常,5%到 10%的代码代表了 99~999%的运行时间。因此,如果你将那 5%的代码卸载到我们的加速器上,那么从技术上讲,你应该能够将应用程序的速度提升 100 倍。这并不是不正常的。我们将图像处理的速度提升 500 倍也并不罕见。现在我们进行数据处理。数据处理是我最喜欢的应用之一,因为几乎所有与机器学习相关的内容都是一种数据驱动的软件方式。数据处理在不断发展。它可以是 SQL 数据处理,也可以是 Spark 类型的数据处理。它可以是向量数据库类型的处理,各种不同的处理方式,无论是非结构化数据还是结构化数据,即数据框。我们极大地加速了这一过程。但为了做到这一点,您必须在其上创建那个库,即动态库。在计算机图形学的情况下,我们很幸运地拥有硅图形作为 OpenGL 和微软的 DirectX。但在这些之外,实际上没有真正的库。因此,例如,我们最著名的库之一是一个库,类似于 SQL,SQL 是一个用于存储计算的库。我们创建了一个名为 cuDNN 的库。cuDNN 是世界上第一个神经网络计算库。因此我们有 cuDNN,我们有 COO,组合优化的选择。我们有 kuQuantum 用于量子模拟和仿真。各种不同的库,例如用于数据处理的数据框的 CDF,SQL。因此,所有这些不同的库都必须被创造出来。这需要将应用程序中运行的算法重构,以便我们的加速器能够运行。如果你使用这些库,那么你将获得 100 倍的提升。


采访人:


Speed up, get much more speed. Incredible.

加速,获得更快的速度。不可思议。


老黄:


And so, so, so the concept is simple and it made a lot of sense. But the problem is how do you go and invent all these algorithms and cause the video game industry to use it?Write these algorithms cause the entire seismic processing and energy industry to use it, write a new algorithm and cause the entire AI industry to use it. You see what I'm saying. And so these libraries, every single one of these libraries, first we had to do the computer science. Second, we have to go through the ecosystem development.And we have to convince everybody to use it. And then what kind of computers doesn't want to run on? All the different computers are different. And so we just did it. One domain after another domain, after another domain, we have, ah, a rich library for self driving cars.We have a fantastic library for robotics, incredible library for virtual screening.Whether it's physics based, virtual screening or neural neural network based virtual screening, incredible library for climatech and so one, one domain after another domain. And so we have to go meet friends and, you know, create the market. And so what Nvidia is really good at, as it turns out, is creating new markets.And we just, we've done it for now, so long that it seems like Nvidia's accelerated computing is everywhere, but we really had to do it one at a time.One industry at a time.

所以,这个概念很简单,而且很有道理。但问题是,如何去发明所有这些算法,并使视频游戏行业使用它?编写这些算法,使整个地震处理和能源行业使用它,编写一个新算法,使整个人工智能行业使用它。你明白我的意思了。因此,这些库,每一个库,首先我们必须进行计算机科学研究。其次,我们必须经过生态系统开发。我们必须说服每个人使用它。那么,什么样的计算机不想运行它呢?所有不同的计算机都是不同的。因此,我们就这样做了,一个领域接着一个领域,我们有,啊,一个丰富的自动驾驶汽车库。我们有一个出色的机器人库,令人难以置信的虚拟筛选库。无论是基于物理的虚拟筛选还是基于神经网络的虚拟筛选,令人难以置信的气候技术库,一个接一个的领域。因此,我们必须去结识朋友,创造市场。因此,事实证明,Nvidia 真正擅长的是创造新市场。我们现在已经做了很长时间,似乎英伟达的加速计算无处不在,但我们真的必须一次一个行业地进行。


采访人:


So I know that many investors in the audience.Are super focused on the data center market, and it would be interesting to kind of get your perspective, the company's perspective on the medium and longterm.Opportunity set obviously your industry is enabling your term, the next industrial revolution. One of the challenges the industry faces is talk a little bit about how you view the data center market as we sit here today.

我知道在座的许多投资者非常关注数据中心市场,了解贵公司的观点以及中长期的机会将会很有趣。显然,您的行业正在推动下一个工业革命。行业面临的挑战之一是,谈谈您如何看待今天我们所处的数据中心市场。


老黄:


There are two things that are happening at the same time and it gets conflated and it's helpful to tease apart.So, the first thing, let's start with

同时发生的有两件事情,它们相互混淆,理清它们是有帮助的。那么,第一件事,我们先从这里开始。


a condition where there's no AI at all.Well, in a world where there's no AI at all, general purpose computing has run out of steam still.And so we know that we know that dinard scaling for all the people in the room that enjoy semiconductor physics, dennard scaling and meetad, Conway's shrinking of transistors, scaling of transistors and dinard scaling of, of a, you know, ISO, ISO power, increased performance or ISO cost, increasing performance that those days are over.And so we're not going to see cpus generalpurpose computers that are going to be twice as fast every year ever again.We'll be lucky if we see it twice as twice as fast every ten years now. Moore's law. Remember, back in the old days, Moore's law was ten times every five years100 times every ten years. And so all we have to do is just wait for the cpus to get faster.And as the world's data centers continue to process more information, cpus got twice as fast every single year. And so we didn't, we didn't see computation inflation. But now that's ended, we're seeing computation inflation. And so the thing that we have to do.Because we have to accelerate everything we can. If you're doing sql processing, accelerate that. If you're doing any kind of data processing at all, accelerate that. If, if you're doing, if you're creating an Internet company and you, you have a recommender system absolutely accelerated and they're now fully accelerated. This a few years ago was all running on cpus but now the world's largest data processing engine, which is a recommender system, is all all accelerated now.And so if you have recommender systems, if you have search systems, any large scale processing of any large amounts of data, you have to just accelerate that. And so, so the first thing that's going to happen is the world's trillion dollars.Of general purpose data centers are going to get modernized into accelerated computing. That's

一种完全没有人工智能的状态。在一个完全没有人工智能的世界里,通用计算已经失去了动力。因此,我们知道,对于在场所有喜欢半导体物理的人来说,Dennard 缩放和 Conway 的晶体管缩小,晶体管的缩放以及 Dennard 缩放,您知道的,ISO 功率、提高性能或 ISO 成本、提高性能的那些日子已经结束。因此,我们不会再看到每年都能快一倍的通用计算机 CPU。现在,如果我们能看到每十年快一倍,那就算幸运了。摩尔定律。记得在过去,摩尔定律是每五年十倍,每十年一百倍。因此,我们所要做的就是等待 CPU 变得更快。随着世界数据中心继续处理更多信息,CPU 每年都快一倍。因此,我们没有看到计算通货膨胀。但现在这一切都结束了,我们看到了计算通货膨胀。因此,我们必须加速一切。如果您在进行 SQL 处理,请加速它。如果您正在进行任何类型的数据处理,请加速这一过程。如果您正在创建一家互联网公司,并且您有一个推荐系统,绝对要加速它,现在它们已经完全加速。几年前,这一切都在 CPU 上运行,但现在世界上最大的 数据处理引擎,即推荐系统,已经完全加速。因此,如果您有推荐系统,如果您有搜索系统,任何大规模处理大量数据的情况,您都必须加速它。因此,首先要发生的事情是,世界上万亿美元的通用数据中心将被现代化为加速计算。那是

going to happen no matter what. That's going to happen no matter what. And, and the reason for that is, as I described, more Las, over and so.The first dynamic you're going to see is the densification of computers.These giant data centers are super inefficient because it's filled with air and air is a lousy conductor of electricity. And so what we want to do is take that few call it.Five thousand, three hundred, two hundred megawatt data center, which is sprawling and you dly it into a really, really small data center. And so if you look at one of our server res, you know, Nvidia server racks look expensive and it could be a couple of million dollars per rack, but it replaces thousands of nodes.The amazing thing is.Just the cables of connecting old general purpose computing systems cost more than replacing all of those and densifying into one rack. The benefit of densifying also is now that you've densified that you can liqu a cool it because it's hard to liquorool a data center that's very large, but you can liqu aool the data center. That's very small.And so the first thing that we're doing is accelerating, modernizing data centers, accelerating it, densifying it, making it more energy efficient. You save money, you save power, you save. And it's much more efficient.That's the first. If we just focused on that, that's.The next ten years, we just.Accelerate that. Now, of course, there's a second dynamic is because of Nvidia's accelerated computing.Brought such enormous cost reductions to computing.It's like in the last ten years, instead of Moore's lobbying 100 eggs, we scaled computing by 1 million X in the last ten years.And so the question is, what would you do different if your plane traveled a million times faster? What would you do different? And so all of a sudden people said, hey, listen, why don't we just use computers to write software?Instead of us trying to figure out what the features are, instead of us trying to figure out what the algorithms are, we'll just give the data, all the data, all the predictive data to the computer and let it figure out what the algorithm is, machine learning, generative AI. And so we did it in such large scale on so many different data domains.That now computers understand, not just.How to process the data but the meaning of the data and because it understands multiple modalities at the same time it can translate data. And so we can go from English to images, images to English, English to proteins, proteins to chemicals. And so because it understood all of the data, the data at one time it can now do all this translation. We call generative AI large amount of text into small amount of text, small amount of text into a large amount of text and so, so on and so forth. We're now in this computer revolution and now what's amazing is.So the first trillion dollars of data centers is going to get accelerated.And invented this new type of software called generative AI. This generative AI is not just a tool, it is a skill.And so this is the interesting thing. This is why a new industry has been created. And the reason for that is if you look at the whole it industry, up until now, we've been making

无论如何都会发生。这将无论如何都会发生。而且,原因如我所描述的,更多的 Las,等等。你将看到的第一个动态是计算机的密集化。这些巨大的数据中心效率极低,因为它们充满了空气,而空气是电的糟糕导体。因此,我们想要做的是将那个五千、三百、二百兆瓦的数据中心,这个庞大的数据中心,缩小成一个非常非常小的数据中心。所以如果你看看我们的服务器机架,你知道,Nvidia 的服务器机架看起来很贵,可能每个机架要几百万美元,但它可以替代成千上万的节点。令人惊讶的是,连接旧的通用计算系统的电缆成本超过了替换所有这些并密集到一个机架的成本。密集化的好处还在于,现在你已经密集化了,你可以进行液体冷却,因为很难对一个非常大的数据中心进行液体冷却,但你可以对一个非常小的数据中心进行液体冷却。因此,我们首先要做的是加速、现代化数据中心,加速它,密集化它,使其更节能。你节省了资金,节省了电力,节省了资源。而且效率更高。这是第一步。如果我们只专注于这一点,那就是接下来的十年,我们只需加速这一点。当然,还有第二个动态,因为 Nvidia 的加速计算为计算带来了巨大的成本降低。在过去的十年里,我们的计算能力不是像摩尔定律所说的每 18 个月翻一番,而是增长了 100 万倍。那么问题是,如果你的飞机旅行速度快了 100 万倍,你会做什么不同的事情?突然间,人们说,嘿,听着,为什么我们不直接用计算机来编写软件呢?我们不再试图弄清楚功能是什么,不再试图弄清楚算法是什么,我们只需将所有数据、所有预测数据提供给计算机,让它来弄清楚算法是什么,机器学习,生成性人工智能。因此,我们在如此大规模的许多不同数据领域中进行了这样的尝试。现在计算机不仅理解。如何处理数据,但数据的意义在于它同时理解多种模式,因此它可以翻译数据。因此,我们可以从英语到图像,从图像到英语,从英语到蛋白质,从蛋白质到化学品。因为它理解了所有数据,所以现在可以进行所有这些翻译。我们称生成性人工智能为将大量文本转化为少量文本,将少量文本转化为大量文本,等等。我们现在正处于这场计算机革命中,令人惊讶的是,第一万亿美元的数据中心将会加速发展。并发明了一种新的软件类型,称为生成性人工智能。这种生成性人工智能不仅仅是一个工具,它是一种技能。因此,这就是有趣的地方。这就是为什么一个新行业被创造出来。并且原因在于,如果你看看整个 IT 行业,直到现在,我们一直在制造

instruments and tools that people use.For the very first time, we're going to create skills that augment people.And so that's why people think that AI is going to expand beyond the trillion dollars of data centers and it and into the world of skills. So what's a skill? A digital chauffeur is a skill autonomous.You know, A A digital assembly line worker, robot, you know, a digital customer service chatbot, digital digital employee for planning, planning, Nvidia supply chain, it could be a, that would be somebody that's a digital SAP agent. There's a, we use a lot of service now in our companies and we have digital employee service. And so now we have all these digital humans essentially.And that's the wave of AI that we're in now, so.

人们使用的工具和仪器。我们将首次创造增强人类的技能。因此,人们认为人工智能将超越万亿美元的数据中心,进入技能的世界。那么,什么是技能?数字司机是一种技能,具有自主性。你知道,数字装配线工人,机器人,数字客户服务聊天机器人,数字规划员工,Nvidia 供应链,这可能是一个数字 SAP 代理。我们在公司中使用了很多 ServiceNow,并且我们有数字员工服务。因此,现在我们基本上拥有所有这些数字人类。这就是我们现在所处的人工智能浪潮。


采访人:


Step back, shift a little. Based on everything you just said, there's definitely an ongoing debate in financial markets.As to whether or not, as we continue to build this AI infrastructure, there is an adequate return on investment.How would you assess customer ROI at this point in the cycle?And if you look back and you kind of think about, you know, PCs cloud computing when they were at similar points in their adoption cycles.How do the rois know look then compared to where we are now as we scale? Yeah, fantastic.

退后一步,稍微调整一下。根据你刚才所说的一切,金融市场中确实存在着持续的辩论。关于在我们继续构建这个人工智能基础设施时,是否有足够的投资回报。你如何评估此时客户的投资回报率?如果你回顾一下,考虑一下个人电脑和云计算在其采用周期的类似阶段时,投资回报率与我们现在在扩展时的情况相比如何?是的,太棒了。


老黄:


So, so let's take a look before cloud. The major trend was virtualization, if you guys remember that. And virtualization basically said, let's take all of the hardware we have in the data center, let's virtualize it into essentially virtual data center, and then we could move workload across the data center instead of associating it directly to a particular computer. As a result, the tendency and the utilization of that data center improved.And we saw essentially a two to one, two point five to one, if you will, cost reduction in data centers, overnight virtualization. The second thing that we then said was after we virtualized that we put that those virtual computers right into the cloud. As a result, multiple companies, not just one companies, many applications, multiple companies can share the same resource, another cost reduction.The utilization again went up.By the way, this last ten years of all of this, fifteen years of all this stuff happening masked the fundamental dynamic which was happening underneath, which is more slow ending.We found a two X, another two X in cost reduction and it hid the end of the transistor, scaling it, hid the transistor, the cpu scaling. Then all of a sudden we already got the utilization cost reductions. Out of both of these things. We're now out and that's the reason why we see.Data center and computing inflation happening right now. And so the first thing that's happening is accelerated computing. And so it's not, it's not uncommon for you to take your data processing work and we if you there's a thing called spark. If you anyone to abuse spark is probably the most used data processing engine in the world today. If you use spark and you accelerate it with Nvidia in the cloud, it's not unusual to see a twenty to one speed up.And so you're going to save ten and you pay. Of course, you got the Nvidia gpu augments the cpu. So the computing cost goes up a little bit. It goes, maybe it doubles, but you reduce the computing time by about twenty times.And so you get a ten X savings, sure. And it's not unusual to see this kind of ROI for accelerated computing. So I would, I would, I would encourage all of you, everything that you can accelerate to accelerate. And then once you accelerated it, run with gpus. And so that's the instant ROI that you get by acceleration.Now beyond that.The generative AI conversation is in the first wave of Gen AI, which is where the infrastructure players like ourselves and all the cloud service providers put the infrastructure in the cloud so that developers could use these machines to train the models and fine tune the models, guardrail the models, so and so forth.And the return on that is fantastic because the demand is so great.For every Dollar that they spend with us translates to five dollars worth of rentals, and that's happening all over the world and everything is all sold out. And so the demand for this is just incredible. Some of the applications that we already know about

所以,让我们在云计算之前先看看。主要趋势是虚拟化,如果你们还记得的话。虚拟化基本上是说,让我们把数据中心所有的硬件虚拟化,基本上变成虚拟数据中心,然后我们可以在数据中心之间移动工作负载,而不是直接与特定计算机关联。因此,数据中心的利用率和趋势得到了改善。我们看到,数据中心的成本减少了大约两倍到两点五倍,虚拟化一夜之间实现。然后我们说的第二件事是,在虚拟化之后,我们将这些虚拟计算机直接放入云中。因此,多个公司,而不仅仅是一家公司,许多应用程序,多个公司可以共享相同的资源,进一步降低成本。利用率再次上升。顺便说一下,在过去的十年、十五年中,所有这些事情的发生掩盖了底层发生的基本动态,那就是更慢的结束。我们发现成本又减少了两倍,这掩盖了晶体管的终结,掩盖了 CPU 的缩放。然后突然之间,我们已经获得了利用成本的降低。基于这两件事。我们现在已经走出这个局面,这就是我们看到数据中心和计算通货膨胀正在发生的原因。因此,首先发生的是加速计算。因此,将数据处理工作进行加速并不罕见。如果你使用一个叫做 Spark 的工具,如果有人滥用 Spark,它可能是当今世界上使用最广泛的数据处理引擎。如果你使用 Spark 并在云中用 Nvidia 进行加速,看到 20 倍的速度提升并不罕见。因此,你将节省 10,而你支付的费用。当然,Nvidia GPU 增强了 CPU。因此,计算成本稍微上升,可能翻倍,但你将计算时间减少了大约 20 倍。因此,你获得了 10 倍的节省,当然。看到这种加速计算的投资回报率并不罕见。所以我会鼓励你们所有人,尽可能加速一切。一旦你加速了,就使用 GPU。因此,这就是你通过加速获得的即时投资回报率。现在,除此之外。生成性人工智能的对话处于生成性人工智能的第一波,这就是像我们这样的基础设施参与者和所有云服务提供商将基础设施放在云端,以便开发者可以使用这些机器来训练模型、微调模型、设置模型的保护措施等等。对此的回报是惊人的,因为需求非常大。他们每花费一美元在我们这里,就能转化为五美元的租赁,这种情况在全球范围内都在发生,一切都已经售罄。因此,对此的需求简直不可思议。我们已经知道的一些应用程序


, of course, the famous ones open AI chatgpt or github copilot or Co generators that we use in our company. The productivity gains are just incredible.There's not one software engineer in our company today who don't use Co generators either. The ones that we build ourselves for kuda or USD which is another language that we use in the company or verilog or c and c plus plus and codegeneration. And so I think the days of every line of code being written by software engineers, those are completely over.And the idea that every one of our software engineers would essentially have companion digital engineers working with them, twenty four, seven, that's the future. And so the way I look at Nvidia, we have 32,000 employees, but those 32,000 employees are surrounded by hopefully 100 X more digital engineers, sure.

当然,著名的有开放 AI 的 ChatGPT 或 GitHub Copilot,或者我们公司使用的 Co 生成器。生产力的提升简直令人难以置信。我们公司今天没有一位软件工程师不使用 Co 生成器。我们为 Kuda 或 USD(这是我们公司使用的另一种语言)或 Verilog 或 C 和 C++以及代码生成自己构建的生成器。因此,我认为每一行代码都由软件工程师编写的时代已经完全结束。我们的每位软件工程师基本上都会有数字工程师伴随他们工作,全天候,这是未来。因此,我看待 Nvidia 的方式是,我们有 32,000 名员工,但这 32,000 名员工周围希望有 100 倍更多的数字工程师,当然。


采访人:


Yes, sure.Lots of industries embracing this.What cases use cases industries are you most excited about?

是的,当然。许多行业正在接受这一点。您最感兴趣的用例行业是什么?


老黄:


Well, in our company, we use it for computer graphics. We can't do computer graphics anymore without artificial intelligence. We compute one pixel, we infer the other 32 just, I mean, it's incredible.And so we hallucinate, if you will, the other 32 and it looks temporarily stable, it looks photorealistic, and the image quality is incredible. The performance is incredible. The amount of energy we save computing one pixel takes a lot of energy. That's computation infncing the other 32 takes very little energy.And you can do it incredibly fast. So one of the takeaways there is AI isn't just about training the model. Of course, that's just the first step. It's about using the model. And so when you use the model, you save enormous amounts of energy, you save enormous amount of time processing time, so we use it for computer graphics. We, if not for AI, wouldn't be able to serve the autonomous vehicle industry if not for AI. The work that we're doing in robotics, digital biology, just about every Tech bio company that I meet these days are built on top of Nvidia. And so they're using it for data processing or generating proteins or for it seems like a.

在我们公司,我们将其用于计算机图形学。没有人工智能,我们无法再进行计算机图形学的工作。我们计算一个像素,推断其他 32 个,我是说,这太不可思议了。因此,我们可以说我们在“幻觉”其他 32 个像素,它看起来暂时稳定,看起来逼真,图像质量令人难以置信。性能也令人难以置信。计算一个像素所节省的能量是巨大的。推断其他 32 个像素所需的能量非常少。而且你可以做到非常快。因此,这里的一个要点是,人工智能不仅仅是训练模型。当然,这只是第一步。关键在于使用模型。因此,当你使用模型时,你节省了大量的能量,节省了大量的处理时间,所以我们将其用于计算机图形学。如果没有人工智能,我们将无法为自动驾驶汽车行业提供服务。我们在机器人技术、数字生物学方面的工作,几乎我最近遇到的每一个科技生物公司都是建立在 Nvidia 之上的。因此,他们将其用于数据处理、生成蛋白质,或者似乎是其他用途。


采访人:


Super exciting.

超级激动人心。


老黄:


It's incredible. Small molecule generation, virtual screening. I mean, just that whole space is going to get reinvented.For the very first time with computer aided drug discovery because of artificial intelligence and so incredible work being done there.

这真是不可思议。小分子生成,虚拟筛选。我的意思是,这整个领域将因计算机辅助药物发现而被重新定义,因为人工智能和在那里所做的令人难以置信的工作。


采访人:


Talk about competition. Talk about your competitive moat. There's certainly group public and private companies looking to disrupt your leadership position. How do you think about your competitive moat?

谈谈竞争。谈谈你的竞争护城河。肯定有一些公共和私人公司希望打破你的领导地位。你如何看待你的竞争护城河?


老黄:


Well, first of all, I think the I would say.I.Several things that are very different about us. The first thing is to remember that AI is not about a chip. AI is about an infrastructure. Today's computing is not build a chip and people come buy your chips, put it into a computer. That's really.Kind of 1990 s.The way that computers are built today, if you look at our new blackwell system, we design seven different types of chips to create the system. Blackwell is one of them.

首先,我想说的是,有几件事让我们非常不同。第一件事是要记住,人工智能并不是关于一个芯片。人工智能是关于基础设施的。今天的计算不是制造一个芯片,然后人们来购买你的芯片,把它放入计算机中。这实际上有点像 1990 年代。今天计算机的构建方式,如果你看看我们的新黑威尔系统,我们设计了七种不同类型的芯片来创建这个系统。黑威尔就是其中之一。


采访人:


And talk about blackwell.

谈谈布莱克威尔。


老黄:


Yeah. And so, yeah. And so the amazing thing is when you, when you, when you want to build this AI computer, people say words like supercluscluster infrastructure supercomputer for good reason.Because it's not a chip, it's not. It's not a computer per se. And so we're building entire data centers by building the entire data center. If you just ever look, look at one of these superclusclusters imagine the software that has to go into it, to write to, to run it. There is no Microsoft Windows for it. Those days are over, so all the software that's inside that computer is completely bespoke. Somebody has to go write that.So the person who designs the chip and the company that designs that, that supercomputer, that superlususter and all the software that goes into it, it makes sense that it's the same company because it'll be more optimized, it'll be more perform it more energy efficient, more cost effective.And so that's the first thing. The second thing is.It is about algorithms and we're really, really good at understanding what is the algorithm, what's the implication to the computing stack underneath and how do I distribute, distribute this computation across millions of processors run for days on days on end with the computer being as resilient as possible, achieving great energy efficiency, getting the job done as fast as possible, so on and so forth. And so we're really, really good at that.And then, lastly, in the end.AI is computing AI software running on computers.And we know that we know that the most important thing for computers is installed base, having the same architecture across every cloud, across onprem to cloud, and having the same architecture available, whether you're building it in the cloud in your own supercomputer, or trying to run it in your car or some robot or some PC that having that same identical architecture that runs all the same software is a big deal. It's called installed base. And so the discipline that we've had for the last thirty years has really LED to today. And it's the reason why the most obvious architecture to use if you were to start a company is to use Nvidia architecture, because we're in every cloud where anywhere you like to buy it.And whatever computer you pick up, so long as it says Nvidia inside, you know you can take the software and run it.

是的。所以,令人惊讶的是,当你想要构建这个人工智能计算机时,人们会说像超级集群基础设施超级计算机这样的词是有原因的。因为这不是一个芯片,也不是一个计算机本身。因此,我们通过构建整个数据中心来构建整个数据中心。如果你曾经看过其中一个超级集群,想象一下必须投入到其中的软件,以便编写和运行它。没有微软 Windows 可以使用,那些日子已经过去了,所以在那台计算机内部的所有软件都是完全定制的。有人必须去编写它。因此,设计芯片的人和设计那个超级计算机、超级集群以及所有投入其中的软件的公司,合理的说应该是同一家公司,因为这样会更优化,性能更好,更节能,更具成本效益。这是第一点。第二点是。这涉及到算法,我们在理解算法、其对底层计算堆栈的影响以及如何将计算分配到数百万个处理器上,持续运行数天方面非常出色,确保计算机尽可能具备弹性,实现出色的能源效率,尽快完成任务,等等。因此,我们在这方面非常出色。最后,人工智能就是在计算机上运行的人工智能软件。我们知道,对于计算机来说,最重要的是安装基础,确保在每个云环境中、从本地到云的架构一致,并且无论是在云中构建自己的超级计算机,还是尝试在汽车、机器人或某台个人电脑上运行,都能拥有相同的架构,运行相同的软件,这一点非常重要。这被称为安装基础。因此,我们在过去三十年中所积累的学科知识确实引领到了今天。这就是为什么如果你要创办一家公司,最明显的架构是使用 Nvidia 架构的原因,因为我们在每个云中,无论你想在哪里购买它。无论你选择什么电脑,只要它标有 Nvidia,你就知道可以运行该软件。


采访人:


Now you're innovating an incredibly fast pace. I want you to talk a little bit more about blackwell 4 X faster on training. Thirty X faster inference than its predecessor hopper. You know, it just seems like you're innovating at such a quick pace. Can you keep up this rapid pace of innovation and.When you think about your partners, how do your partners keep up with the pace of innovation you're delivering?

现在你们的创新速度非常快。我希望你能多谈谈黑威尔(Blackwell)在训练上快了 4 倍,推理速度比前身霍普(Hopper)快了 30 倍。你知道,这似乎你们的创新速度如此之快。你们能否保持这种快速的创新步伐?当你考虑到你们的合作伙伴时,他们是如何跟上你们所提供的创新步伐的?


老黄:


The pace of innovation, our basic methodology is to take, because remember we're building an infrastructure. There are seven different chips.Each chips.Rhythm is probably, at best, two years, at best, two years.We could give it a midlife kicker every year, but architecturally, if you're coming up with a new architecture every two years, you're running at the speed of light. Okay? You're running insanely fast.Now we have seven different chips and they all contribute to the performance. And so we could innovate and bring a new AI cluster, a superclus to the market every single year. That's better than the last generation because we have so many different pieces to work around and so.And the benefit of performance at the scale that we're doing, it directly translates the t, and so when blackwell is three X the performance.For somebody who has a given amount of power, say one gigawatt.That's three times more revenues.That performance translates to throughput, that throughput translates to revenues. And so for somebody who has a gigawatt of power to use, you get three X the revenues.There's no way you can give somebody a cost reduction or a discount on chips to make up for three X the revenues. And so the ability for us to deliver that much more performance through the integration of all these different parts and optimizing across the whole stack and optimizing across the whole cluster, we can now deliver better and better value at much higher rates.The opposite of that is equally true.For any amount of money you want to spend, so for ISO power you get three X the revenues for ISO.Spend you get three X the performance, which is another way of saying cost reduction. And so we have the best per per wide, which is your revenues. We have the best per per t, which means your gross margins and so we keep pushing this out to the marketplace. Customers get to benefit from that not once every two years and it's architecturally compatible. And so the software you developed yesterday will run tomorrow. The software you developed today will run across your entire installed base so we could run incredibly fast.If every single architecture was different, then you can do this.Takes a year just to cobble together a system because we built everything together the day we ship it to you and it's pretty famous. Somebody tweeted out that in nineteen days after we shipped systems to them, they had a super cluster up and running. Nineteen days.You can't do that if you are cobbling together all these different chips and writing the software. You'll be lucky if you could do it in a year. And so I think our ability to transfer our innovation pace.To customers getting more revenues, getting better gross margins, that's a fantastic thing.

创新的步伐,我们的基本方法是采取,因为请记住我们正在构建基础设施。有七种不同的芯片。每种芯片的节奏可能,最多,两年,最多,两年。我们可以每年给它一个中期提升,但从架构上讲,如果你每两年推出一个新架构,你就是在以光速运行。好吧?你在疯狂地快速前进。现在我们有七种不同的芯片,它们都对性能有所贡献。因此,我们可以每年创新并推出一个新的 AI 集群,一个超级集群。这比上一代更好,因为我们有这么多不同的组件可以使用。而且我们所做的规模上的性能收益,直接转化为 t,因此当 blackwell 的性能是三倍时。对于某个拥有一定功率的人,比如一吉瓦。这是三倍的收入。该性能转化为吞吐量,该吞吐量转化为收入。因此,对于某个有一吉瓦功率可用的人来说,你可以获得三倍的收入。没有办法通过降低芯片成本或提供折扣来弥补三倍的收入。因此,我们通过整合所有这些不同的部分并在整个堆栈和整个集群中进行优化,能够提供更多的性能,现在可以以更高的速度提供更好的价值。相反的情况同样成立。无论你想花多少钱,对于 ISO 功率,你可以获得三倍的 ISO 收入。花费可以获得三倍的性能,这也是另一种说法,意味着成本降低。因此,我们在每单位收入方面表现最佳,我们在每单位毛利方面表现最佳,因此我们不断将这一点推向市场。客户从中受益,不是每两年一次,而且在架构上是兼容的。因此,你昨天开发的软件明天可以运行。你今天开发的软件可以在你整个已安装基础上运行,这样我们就可以运行得非常快。如果每个架构都不同,那么你就无法做到这一点。花费一年时间仅仅是为了拼凑一个系统,因为我们在发货当天将所有东西都整合在一起,这一点相当有名。有人在推特上发文称,在我们向他们发货后的十九天内,他们就已经搭建好了超级集群。十九天。如果你还在拼凑这些不同的芯片并编写软件,那是做不到的。你能在一年内做到就算幸运了。因此,我认为我们将创新速度转移给客户的能力,帮助他们获得更多收入、提高毛利率,这是一件了不起的事情。


采访人:


The majority of your supply chain partners operate out of Asia, particularly Taiwan. Given what's going on geopolitically, how you're thinking about that as you look forward?

大多数供应链合作伙伴在亚洲运营,特别是台湾。考虑到当前的地缘政治形势,您在展望未来时是如何看待这一点的?


老黄:


The, the Asia supply chain, as you know, is really, really sprawling and interconnected.People think that when we say gpus because a long time ago.When I announced a new chip, a new generation of chips, I would hold up the chip and so that was a new gpu Nvidia new gpus are 35,000 parts, weighs eighty pounds.You know, consumes 10,000 amps.When you rack it up, it weighs 3000 pounds' and so the, these gpus are so complex. It's built like a like electric car components, like an electric car. And so the ecosystem is really diverse and really interconnected in Asia.We try to.Design diversity and redundancy into every aspect wherever we can.And then

亚洲供应链,如你所知,确实非常庞大且相互关联。人们认为,当我们提到 GPU 时,是因为很久以前。当我宣布一款新芯片,一代新芯片时,我会举起芯片,所以那是一个新的 GPU。Nvidia 的新 GPU 有 35,000 个部件,重达 80 磅。你知道,消耗 10,000 安培。当你把它们放在机架上时,重达 3000 磅。这些 GPU 是如此复杂。它的构造就像电动车的组件,像一辆电动车。因此,生态系统在亚洲确实非常多样化且相互关联。我们尽量在每个方面设计多样性和冗余。然后


the last part of it is to have enough intellectual property in our company.In the event that we have to shift from one fab to another, we have the ability to do it. Maybe the process technology is not as great. Maybe, you know, we won't be able to get the same level of of, of performance or cost, but we will be able to provide the supply. And so, so I think the in the event anything were to happen we should be able to pick up and fab it somewhere else. We're fabbing at a TSMC because it's the world's best.And it's the world's best, not by a small margin as the world. It's about a lot by this incredible margin. And so not only just the long history of working with them, the great chemistry, their agility, the fact that they could scale, you know, remember Nvidia, last year's revenue had a major hockey stick.That major hockey stick wouldn't have been possible if not for the supply chain responding. And so the agility of that supply chain, including TSMC, is incredible. And in just less than a year, we've scaled up coos capacity tremendously. And we're going to have to scale it up even more next year and scale up even more the year after that. But nonetheless, the agility and their capability to respond to our needs is just incredible. And so we use them because they're great, but if necessary, of course, we can always bring up others.

最后一部分是确保我们公司拥有足够的知识产权。如果我们必须从一个晶圆厂转移到另一个晶圆厂,我们有能力做到这一点。也许工艺技术不是那么出色。也许,你知道,我们无法获得相同水平的性能或成本,但我们能够提供供应。因此,我认为如果发生任何事情,我们应该能够在其他地方进行生产。我们在台积电生产,因为它是世界上最好的。而且它是世界上最好的,差距并不小,而是非常大的差距。因此,不仅仅是与他们长期合作的历史,良好的化学反应,他们的灵活性,以及他们能够扩展的事实,你知道,记得英伟达,去年的收入有一个巨大的曲线增长。如果没有供应链的响应,这个巨大的曲线增长是不可能的。因此,包括台积电在内的供应链的灵活性是惊人的。在不到一年的时间里,我们的产能大幅提升。明年我们还需要进一步扩大产能,后年还要继续扩大。但尽管如此,他们的灵活性和响应我们需求的能力真是令人难以置信。因此,我们使用他们,因为他们很出色,但如果有必要,当然我们可以随时引入其他人。


采访人:


Company incredibly well positioned.Lot of great stuff we've talked about. What do you worry about?

公司位置极其优越。我们谈论了很多好东西。你担心什么?


老黄:


Um.Well.I.Our company.Works with every AI company in the world today.We're working with every single data center in the world today.I don't know one data center, one cloud service provider, one computer maker we're not working with.And so what comes with that is enormous responsibility. And we have a lot of people on our shoulders and everybody's counting on us. And demand is so great that that delivery of our components and our technology and our infrastructure and software is really emotional for people because it directly affects our revenues, it directly affects our competitiveness.And so we probably have more emotional customers today than andervedly. So if we could fulfill everybody's needs, then the emotion would go away. But it's very emotional. It's really tense. We've got a lot of responsibility on our shoulder and we're trying to do the best we can. And here we are ramping blackwell and it's in full production. We'll ship in Q4 and scale it, start scaling in Q4 and into next year and, and the.The demand on it is so great and everybody wants to be first and everybody wants to be most and everybody wants to be. And so the intensity is really, really quite extraordinary. And so I think it's fun to be. It's fun to be inventing the next computer era. It's fun to see all these amazing applications being created. It's incredible to see robots walking around.You know, it's incredible to have, have these digital agents coming together as a team, solving problems in your computer. It's amazing to see theis that we're using to design the chips that will run ouris.All of that stuff is incredible to see. The part of it that is just really intense is just the world on our shoulders and so sure, so less sleep is fine and we're.Three solid hours. That's all we need well.

嗯。好吧。我。我们公司。与今天世界上每一个人工智能公司合作。我们正在与世界上每一个数据中心合作。我不知道有哪个数据中心、哪个云服务提供商、哪个计算机制造商我们没有合作。因此,这带来了巨大的责任。我们肩上承载着很多人的期望,每个人都在依赖我们。需求如此之大,以至于我们组件、技术、基础设施和软件的交付对人们来说真的很重要,因为这直接影响我们的收入,直接影响我们的竞争力。因此,我们今天可能有更多情绪化的客户。如果我们能够满足每个人的需求,那么情绪就会消失。但这真的很情绪化。气氛非常紧张。我们肩上有很大的责任,我们正在尽力而为。现在我们正在加速黑威尔的生产,并且已经全面生产。我们将在第四季度发货,并开始在第四季度及明年进行扩展。对它的需求非常大,每个人都想成为第一,每个人都想拥有最多,每个人都想要。因此,紧迫感真的非常、非常强烈。所以我觉得这很有趣。发明下一个计算机时代很有趣。看到这些惊人的应用程序被创造出来很有趣。看到机器人四处走动真是不可思议。你知道,看到这些数字代理人作为一个团队聚集在一起,解决你计算机中的问题是令人难以置信的。看到我们用来设计将运行我们的芯片的东西真是令人惊叹。所有这些东西都令人难以置信。真正紧张的部分就是肩上的重担,所以,当然,少睡觉也没关系,我们只需要三小时的高质量睡眠。


采访人:


Good for you. I need more than that. I could, I could spend another half hour. Unfortunately, we've got to stop. Jensen, thank you very much. Thank you for being you, Chairman.

对你来说很好。我需要的不止这些。我可以,我可以再花半个小时。不幸的是,我们必须停止。詹森,非常感谢你。谢谢你做你自己,主席。


老黄:


Thank you. Thank you.

谢谢。谢谢。



180K
Knowledge builds up, like compound interest
 最新文章