Name: NVIDIA執行長黃仁勳 Microsoft Build 2025 爐邊對談！ (Conversation with NVIDIA CEO Jensen Huang: Satya Nadella at Microsoft Build 2025)
Uploaded: 2025-12-03T07:11:07.000Z
Duration: 12 min 36 s

Thank you so much, Jensen, for joining yet again for our Build Developer Conference.

謝謝 Jensen 再次來到我們的 Build 開發者大會。

You were here a couple of years ago and we had a chance at that time to talk about all the great innovation you are bringing to our Azure and how customers were going to benefit.

幾年前你來過，那時候我們有機會聊到你為我們的 Azure 帶來了多少很棒的創新，以及客戶將如何受益。

And in fact, in the last two years you have sort of continued to push the real frontier of innovation.

事實上，在過去兩年裡，你持續推動著創新的真正前沿。

And you and I have talked a lot about how this may be again the golden age of both Silicon hardware systems, software systems, all coming together.

你和我聊了很多，這可能又會是矽晶片硬體系統、軟體系統，所有東西結合在一起的黃金時代。

And one of the things that I've really learned so much from you in sort of saying, look, at the end of the day, our joint goal is to be able to deliver more intelligence to the world.

其中一件事是我從你身上學到很多的，就是說，歸根究柢，我們的共同目標是能夠為世界帶來更多智慧。

In some sense, you can even say it's tokens per dollar per watt, right?

在某種程度上，你甚至可以說，這是每瓦每美元的代幣數量，對吧？

That's sort of what ultimately is going to help the world flourish.

這就是最終將幫助世界蓬勃發展的東西。

And so I just wanted to be able to maybe get your perspective on all this, maybe starting in fact with the very start of what has been our, which is Moore's Law.

所以我想聽聽你對這一切的看法，也許可以從我們一直以來的起點開始，也就是摩爾定律。

You're really taken this Moore's Law and put it back on hyperdrive and maybe you want to speak to.

你真的把這個摩爾定律重新帶回了超速模式，也許你想談談。

In fact, right behind me is a signed GB200 from you and we are now just about putting that into massive production.

事實上，在我身後就是你簽名的 GB200，我們現在正準備將它投入大規模生產。

And so I just want you to talk a little bit about how this is actually working generation to generation in terms of just the benefits of Moore's Law.

所以我想請你稍微談談，從這一代到下一代，摩爾定律的好處究竟是如何實現的。

In fact, two years ago we had just launched the largest AI supercomputer in the world together on Azure.

事實上，兩年前我們才剛在 Azure 上推出了當時全球最大的 AI 超級電腦。

And then right now, as we speak, we are in full production with Grace Blackwell.

而現在，就在我們談話的同時，Grace Blackwell 已經進入全面量產。

We are ramping and scaling and building the largest AI supercomputer in the world in Azure.

我們正在 Azure 上加速擴大規模，建造全球最大的 AI 超級電腦。

So this is just an exciting, exciting time literally two years later.

所以，這真的是一個令人興奮的時刻，就在短短兩年後。

Well, during that time we've innovated across the entire stack.

嗯，在這段期間，我們在整個堆疊上都進行了創新。

This is the big change in how computing works now.

這就是現在運算方式的重大改變。

Instead of just the processor running static software, if you will, everything about the stack has changed.

不再只是處理器運行靜態軟體，可以說，堆疊的每個部分都改變了。

And so our processor has changed, our MV link has changed so that we can scale up to a much larger compute node.

所以我們的處理器改變了，我們的 MV link 也改變了，這樣我們才能擴展到更大的運算節點。

It's liquid cooled, the architecture is FP4 tensor core.

它是液體冷卻的，架構是 FP4 tensor core。

We now have the ability to connect Grace and Blackwell coherently over a super fast link.

我們現在能夠透過超高速連結，將 Grace 和 Blackwell 連接在一起。

Very important in the world of KV caching.

這在 KV caching 的世界裡非常重要。

As we talk about these large agentic model and agentic workforces, workloads, all of this stuff combined with our new CUDA algorithms and new model technology in your new AI infrastructure on Azure together, all of that together 40x speed up over Hopper 40x speed up over Hopper.

當我們談論這些大型的 agentic model 和 agentic workforce、workloads，所有這些東西結合我們新的 CUDA 演算法和新的 model 技術，在你的 Azure 新 AI 基礎設施上，所有這些加起來比 Hopper 快了 40 倍，比 Hopper 快了 40 倍。

That's just an insane speed up in just two years.

這在短短兩年內就達到了驚人的加速。

And in fact, you know that type of compounding of S curves, right?

事實上，你知道那種 S 型曲線的複利效應，對吧？

Your innovation, our innovation coming together, ultimately driving that frontier forward is just unbelievable.

你的創新，我們的創新結合在一起，最終推動了前沿的發展，這真是令人難以置信。

But one of the other things you taught me, which I'll always remember, is that speed of light execution between even the two organizations is a way to deliver that right, which is there is in fact, you know, in our fleet, as you can imagine, we have all the generations of Nvidia and the new generations coming in, and

但你教我的另一件事，我永遠不會忘記，那就是即使是兩個組織之間的光速執行，也是一種交付方式，對吧？事實上，你可以想像，在我們的機隊裡，我們有所有世代的 Nvidia 和新世代的產品，

you want to speak to a little bit of what Moore's Law and speed have to do with each other when it comes to driving innovation.

當談到推動創新時，你想談談摩爾定律和速度之間的關係。

Yeah, this is the insane execution of our two organizations now, hyperlinked, if you will.

對啊，這就是我們兩個組織現在令人難以置信的執行力，如果可以這麼說的話，是緊密連結在一起的。

In fact, when technology is moving 40 times per generation, and

事實上，當科技每一代進步 40 倍時，

it's 40 times every two years, you really want to upgrade every year.

而且是每兩年進步 40 倍，你真的會想每年都升級。

You don't want to wait every four years, build out a giant fleet.

你不想等四年才一次性建置龐大的設備。

You want to build out small amounts at every single year.

And each year, let's say it's 10 times faster than the last year.

而且假設每一年都比前一年快 10 倍。

And as a result, you, you know, cost average, if you will, or inverse to that performance, average up your entire fleet.

結果就是，你，你知道，平均成本，或者反過來說，平均效能，讓你的整個設備都提升。

And so you want to, you want to innovate and you want to integrate new data centers, new computer architectures every year.

所以你想，你想創新，而且你想每年整合新的資料中心、新的電腦架構。

The complexity, of course, is the work that you and I do is not about building PCs anymore.

當然，複雜性在於，你和我做的不再是建置個人電腦。

And for our organizations now to be innovating every single year requires really type integration.

我們的組織現在要做到每年創新，真的需要深度整合。

Yeah, and I think that that point about that physics of the fleet, right.

對，我認為關於艦隊的物理學這一點，對吧。

Continuously taking advantage of what comes out each year, and

持續利用每年推出的新技術，而且

that compounding of that 40x improvement every two years, let's say, and

then the fleet itself will be there for four years, five years.

然後整個艦隊會運作四年、五年。

And the benefits of it go all through that period.

在這整個期間，它的好處都會持續存在。

In fact, talking about that, one of the other things you mentioned this even in, you know, when you talked about the first point, which is the software, the innovation you're doing in Cuda, some of the system software work that we do end to end, the app, we talked a lot about, even the new app server for AI that we are doing even something like, even just, you know, when I look at an application and its performance, whether it's latency cogs, even the smart techniques of prompt caching, plus everything else that happens Just makes a huge difference.

事實上，說到這個，你還提到了一件事，你知道，當你談到第一點，也就是軟體，你在 Cuda 裡做的創新，我們端到端的系統軟體工作，應用程式，我們談了很多，甚至我們正在做的新的 AI 應用程式伺服器，就像，你知道，當我看一個應用程式的效能時，無論是延遲、核心，甚至是 prompt caching 的智慧技術，加上其他所有發生的事情，都會帶來巨大的差異。

But you want to talk about the compounding effects of software that then add to all of what we are doing together.

但你想談談軟體的複利效應，然後為我們共同做的一切增添價值。

這在運算方面真是太不可思議了。

Architectures that are stable, that are programmable, have a rich ecosystem.

架構穩定、可程式化，並擁有豐富的生態系統。

So meanwhile, while we're sitting here executing from Pascal to Ampere to Hopper to Blackwell to Next Generation, all of our software architectures are compatible.

所以，當我們從 Pascal、Ampere、Hopper、Blackwell 到下一代產品不斷推進時，我們所有的軟體架構都是相容的。

As a result, that rich ecosystem of tools and all the developers that are developing on it, they're anxious to develop on it because the install base is so large, they're anxious to develop on it because the ecosystem is so rich.

是以，這個豐富的工具生態系統以及所有在其上開發的開發者，他們渴望在其上開發，因為安裝基礎龐大，他們渴望在其上開發，因為生態系統非常豐富。

And all of that enables them to invest deeply into enhancing their models and their algorithms and runtimes and orchestration layers because it affects and benefits the entire fleet.

這一切都讓他們能夠深入投資於增強他們的模型、演算法、執行階段和協調層，因為這會影響並惠及整個艦隊。

And so while we're sitting here executing new generations of architectures every single year, we still want architecture compatibility.

所以，當我們每年不斷推出新一代架構時，我們仍然希望架構相容。

We want the rich ecosystem to be stable so that software developers investments, and

我們希望豐富的生態系統保持穩定，以便軟體開發者的投資，以及

all of the AI developers investments and all your infrastructure customers investments get to be distributed and amortized across the entire fleet.

所有 AI 開發者的投資和所有基礎設施客戶的投資都能在整個艦隊中分攤和攤銷。

in fact, you know, one of the things we're very excited about is given the fleet we now have, we are able to take advantage of these software advances across the entire fleet.

事實上，你知道，我們很興奮的一件事是，考慮到我們現在擁有的整個艦隊，我們能夠在整個艦隊中利用這些軟體進步。

And I think that's one of the things that people don't understand, that even the things that were shipped from you multiple years ago benefit from software advances continuously.

我認為這也是大家不太瞭解的一點，即使是你們好幾年前出貨的產品，也能持續受益於軟體進步。

Hopper, we've been shipping Hoppers to you now for two years.

Hopper，我們出貨 Hopper 給你們已經兩年了。

Over the course of those two years, through new algorithms, let's just say transformer, let's just say llama 70B.

在這兩年中，透過新的演算法，像是 transformer，像是 llama 70B。

Over the course of the last two years, we've improved the performance of 40x Blackwall to Hopper is 40x.

在過去兩年中，我們將 40x Blackwall 的效能提升到 Hopper 的 40x。

Hopper on Hopper has been 40x over the last two years.

Hopper 在 Hopper 上過去兩年已經提升了 40x。

We did that through all kinds of things like in flight batching and speculative decoding and all kinds of new algorithms that sit underneath the inference engine.

我們透過了許多方式達成這點，像是即時批次處理 (in-flight batching) 和推測解碼 (speculative decoding)，以及推理引擎底層的各種新演算法。

of course, all of the researchers and all the developers, they want to invest in improving the architecture on CUDA because they know the install base is so large, whatever work that they do will benefit a lot of the fleet and many users will be able to benefit from it.

當然，所有的研究人員和開發者，他們都想投資於改善 CUDA 上的架構，因為他們知道安裝基礎非常龐大，他們所做的任何工作都會讓整個艦隊受益，許多使用者也能從中獲益。

And so you're willing to dedicate your entire, you know, all of your heart and soul into optimizing the algorithm and the architecture because the installed base is so large.

所以你願意全心全意地投入演算法和架構的優化，因為現有的基礎非常龐大。

That's so well said because in some sense, you know, the latest GB200 coming up means some of the folks who are doing the cutting Edge new training will go there.

說得太好了，因為某種程度上，你知道，最新的 GB200 即將推出，代表一些走在最前沿進行新訓練的人會轉移到那裡。

And at the same time we're also providing, for example, an Azure container service, some of the sort of software gains of the fleet on the GPU fleet so that they can run any agent on it.

同時我們也提供，例如 Azure container service，讓 GPU 艦隊的一些軟體優勢，能夠在上面運行任何代理程式。

So that ability for us to be able to flex the fleet for the various different parts of the frontier of latency cogs and performance is just fantastic to have.

所以我們能夠為各種前沿的延遲和效能需求靈活運用艦隊的能力，真是太棒了。

Yeah, your engineers and ours are at this moment optimizing runtimes on Ampere A10s, A1 hundreds.

沒錯，你們和我們的工程師們此刻正在 Ampere A10、A100 上優化運行時間。

And the beautiful thing is because of our natural understanding of the importance of preserving enhancing developer productivity and developer value for the life of the architecture, we support software and we fine tune software for as long as we shall live.

而最棒的是，因為我們自然而然地理解在架構的整個生命週期中，維護和提升開發者生產力與價值的importance，我們會一直支援軟體並微調軟體。

And all of that is just between our two companies.

而這一切都僅限於我們兩家公司之間。

And then the last one I wanted to unpack is another thing you talk about sort of GPUs is sort of more broader than just one AI workload.

然後最後一個我想深入探討的，是你們也提過的，關於 GPU 的部分，它其實比單一的 AI 工作負載更廣泛。

It is this accelerated compute vision you've had, quite frankly for, you know, the longest time.

這就是你們一直以來，說實話，長期以來都在推動的加速運算願景。

But I think that also is a massive benefit to us because that also means we can bring many, many different classes of workloads.

但我認為這對我們來說也是一個巨大的優勢，因為這也意味著我們可以導入許多、許多不同類型的應用。

That means customers can use this fleet for of course the cutting edge AI, but many workloads.

這表示客戶當然可以用這套運算資源來處理最尖端的 AI，也能處理很多其他的應用。

But you want to speak a little bit to the what happens to UTIL and the diversity of workloads that get all accelerated.

但您想稍微談談，關於 UTIL 的情況，以及有多少不同類型的應用都能獲得加速嗎？

One of the benefits of CUDA is that the installed base is so large.

CUDA 的一個好處是，它的安裝基礎非常龐大。

Another benefit of CUDA is that on the one hand it's an accelerator architecture, on the other hand it's fairly general purpose for some of the heaviest workloads.

CUDA 的另一個好處是，一方面它是加速器架構，另一方面，對於一些最繁重的應用，它的通用性也很高。

Our two teams are working on accelerating data processing, accelerating data processing by 20x50x, transcoding video image processing models of all kinds, recommender systems, vector search engines, all of those different types of algorithms map wonderfully on top of cuda.

我們的兩個團隊正致力於加速資料處理，將資料處理速度提升 20 倍到 50 倍，還有各種影片轉碼、影像處理模型、推薦系統、向量搜尋引擎，所有這些不同類型的演算法都能完美地運行在 CUDA 上。

And so with our computer scientists working together, we can accelerate one workload after another workload, one after another workload.

所以，透過我們的電腦科學家們一起努力，我們可以一個接一個地加速工作負載。

And so as this tail, if you will, the long tail of GPUs and all of the frontier models move off to Blackwells and futures, we leave behind this compute fleet which is still orders of magnitude faster than general purpose computing that can be used for all these different workloads.

是以，隨著這條長長的 GPU 產品線，以及所有最前沿的模型轉移到 Blackwell 和未來的產品上，我們留下了這支運算資源，它仍然比通用運算快上幾個數量級，可以用於所有這些不同的應用。

And so our job is to make sure that that fleet and that workload, all of the workloads of your data center are fully accelerated.

所以我們的任務就是確保這整個艦隊和工作負載，還有你資料中心的所有工作負載都能被完全加速。

The fleet is fully utilized for the entire life of the fleet.

讓整個艦隊在它的整個生命週期內都能被充分利用。

And so wealth captures that real essence.

所以，財富捕捉到了這個真正的精髓。

Ultimately, it's not just tokens, but it's all workloads per dollar, per watt.

最終，不僅是 token，而是每瓦每美元的所有工作負載。

Can we really accelerate them all as we continue to innovate across both the hardware and the software boundaries and really have the compounding effects ultimately show up for our customers?

隨著我們持續在硬體和軟體邊界上創新，我們真的能加速它們嗎？並讓這些複利效應最終呈現在我們的客戶身上嗎？

So thank you so much, Jensen, for the partnership.

所以，Jensen，非常感謝你的合作。

Thanks to everybody at Nvidia for the amazing work that you all are doing.

感謝 Nvidia 的所有人，感謝你們所做的驚人工作。

字幕列表影片播放

NVIDIA執行長黃仁勳 Microsoft Build 2025 爐邊對談！ (Conversation with NVIDIA CEO Jensen Huang: Satya Nadella at Microsoft Build 2025)

ultimately

sort

entire

incredible