賈氏柏的屋簷下

黃仁勳 GTC Taipei 2026 演講搞

2026-06-01T22:02:38.485+08:00

00:08:42,000 --> 00:08:48,000

So many. Taiwan's rich ecosystem.

這麼多。台灣的生態系統真豐富。

00:08:48,000 --> 00:08:53,000

The richest ecosystem. The world's best supply chain ecosystem.

最豐富的生態系統。全球最佳的供應鏈生態系統。

00:08:53,000 --> 00:08:56,000

Unbelievable. Well, thank you all for being here.

難以置信。好，感謝各位蒞臨。

00:08:56,000 --> 00:09:02,000

And this year, this year, our businesses together are growing incredibly.

而今年，今年，我們共同的業務成長令人難以置信。

00:09:02,000 --> 00:09:11,000

In fact, somebody told me last night that the annual GDP of Taiwan is going to grow almost 10%.

事實上，昨晚有人告訴我，台灣的年度 GDP 即將成長近 10%。

00:09:17,000 --> 00:09:22,000

Unbelievable. Well, we have a lot to talk about. Let's get going.

難以置信。好了，我們有很多要聊的。開始吧。

00:09:22,000 --> 00:09:27,000

Two years ago when I was here, I started to talk to you about how AI has moved from

兩年前我在這裡時，開始和你們談論 AI 如何從

00:09:27,000 --> 00:09:31,000

generative AI and the other waves of AIs that are coming.

生成式 AI 演進，以及其他即將到來的 AI 浪潮。

00:09:31,000 --> 00:09:34,000

The next wave of AI was agentic AI.

AI 的下一波浪潮是 Agentic AI（自主代理 AI）。

00:09:34,000 --> 00:09:41,000

And today, we can say that agentic AI has arrived, that useful AI has arrived.

而今天，我們可以宣告：Agentic AI 已經到來，真正有用的 AI 已經到來。

00:09:41,000 --> 00:09:44,000

Now, what does this mean? This is GitHub.

這意味著什麼？這是 GitHub。

00:09:44,000 --> 00:09:50,000

This is, of course, one of the first applications of agentic AI is software coding.

當然，Agentic AI 最早的應用之一就是軟體編程。

00:09:50,000 --> 00:09:53,000

One of the most valuable professions.

這是最有價值的職業之一。

00:09:53,000 --> 00:09:55,000

Incredibly large ecosystem.

規模龐大、令人難以置信的生態系統。

00:09:55,000 --> 00:10:00,000

30 million, 40 million professional software developers.

3 千萬到 4 千萬名專業軟體開發者。

00:10:00,000 --> 00:10:06,000

Probably another couple of hundred who are students and enthusiasts and so on and so forth.

還有大概幾億名學生和愛好者等等。

00:10:06,000 --> 00:10:12,000

But say 30, 40 million software developers in the world code for a living.

但粗算全球有 3 到 4 千萬名靠編程維生的軟體開發者。

00:10:12,000 --> 00:10:15,000

And this represents most of them.

而這代表了其中的大多數。

00:10:15,000 --> 00:10:23,000

This is GitHub. The pull request is when they download software, they modify it, and commit is when they push it back up.

這是 GitHub。Pull request 是下載軟體並修改，commit 是將修改推送回去。

00:10:23,000 --> 00:10:34,000

Okay? And so if you could look at this, in 2023, the number of commits was 300 million.

好的？如果你看這個數據，2023 年的 commit 次數是 3 億次。

00:10:34,000 --> 00:10:37,000

2024, 400 million.

2024 年，4 億次。

00:10:37,000 --> 00:10:43,000

2025, 500 million commits.

2025 年，5 億次 commit。

00:10:43,000 --> 00:10:51,000

In the first few months, in the first few months of 2026, it has nearly tripled.

在 2026 年的前幾個月，前幾個月，已幾乎成長了三倍。

00:10:51,000 --> 00:10:54,000

Now, what does that mean?

這代表什麼意思？

00:10:54,000 --> 00:11:02,000

30 million software developers representing about $3 trillion worth of GDP.

3 千萬名軟體開發者，代表約 3 兆美元的 GDP。

00:11:02,000 --> 00:11:08,000

Producing three, that's what they're paid, $3 trillion worth of salaries per year,

創造了——那是他們的薪資——每年 3 兆美元的薪資，

00:11:08,000 --> 00:11:14,000

which is generating economic growth for the rest of the industries.

這推動了其他所有產業的經濟成長。

00:11:14,000 --> 00:11:23,000

Say $100 trillion of the world's industries is impacted, is generated by $3 billion worth of salary.

也就是說，世界上約 100 兆美元的產業，是由這 3 兆美元薪資的軟體工程師所驅動的。

00:11:23,000 --> 00:11:34,000

That $3 trillion, excuse me, $3 trillion, that $3 trillion worth of salary is now producing nearly three times as much output.

那 3 兆美元——不好意思——那 3 兆美元，那 3 兆美元的薪資現在正在創造將近三倍的產出。

00:11:34,000 --> 00:11:43,000

It's effectively a $9 trillion productivity from $3 trillion of salaries.

實際上，3 兆美元的薪資創造出了 9 兆美元的生產力。

00:11:43,000 --> 00:11:45,000

Does that make any sense?

這有道理嗎？

00:11:45,000 --> 00:11:48,000

The difference is absolutely extraordinary.

這個差距是絕對驚人的。

00:11:48,000 --> 00:11:50,000

This is the potential. This is the promise of AI.

這就是 AI 的潛力。這就是 AI 的承諾。

00:11:50,000 --> 00:11:54,000

The number of engineers, software engineers, is actually increasing.

工程師、軟體工程師的數量，實際上正在增加。

00:11:54,000 --> 00:11:58,000

People talk about AI reducing jobs, complete nonsense.

人們說 AI 會減少工作，這完全是無稽之談。

00:11:58,000 --> 00:12:02,000

It's causing more software engineers to be hired.

AI 反而促使更多軟體工程師被雇用。

00:12:02,000 --> 00:12:04,000

The reason for that is very simple.

原因非常簡單。

00:12:04,000 --> 00:12:12,000

If you can hire a software engineer and you could generate $9 trillion worth of productive work,

如果你能雇用一名軟體工程師，並能產生 9 兆美元的生產性工作，

00:12:12,000 --> 00:12:15,000

why wouldn't you want to hire more software engineers?

你怎麼會不想雇用更多軟體工程師呢？

00:12:15,000 --> 00:12:22,000

If that line was flat, then obviously people will hire fewer software engineers.

如果那條線是持平的，那顯然人們會雇用較少的軟體工程師。

00:12:22,000 --> 00:12:26,000

But because the output is so incredible, people want to hire more software engineers.

但正因為產出如此驚人，人們希望雇用更多軟體工程師。

00:12:26,000 --> 00:12:30,000

This is going to show up in our economy somehow soon.

這很快就會以某種形式在我們的經濟中顯現。

00:12:30,000 --> 00:12:33,000

So the first thing is useful AI has arrived.

所以第一件事是：真正有用的 AI 已經到來。

00:12:33,000 --> 00:12:36,000

Now, what does that mean from the industry's perspective?

從產業角度來看，這意味著什麼？

00:12:36,000 --> 00:12:42,000

From the industry's perspective, that means that tokens are now in extraordinary demand.

從產業角度來看，這意味著 token 現在面臨極大的需求。

00:12:42,000 --> 00:12:45,000

Because if you could do this, you're going to want to produce more of it.

因為如果你能做到這一點，你就會想生產更多。

00:12:45,000 --> 00:12:54,000

And because tokens are now profitable units, tokens are now profitable units of revenues.

而因為 token 現在是有利可圖的單位，token 現在是有利可圖的營收單位。

00:12:54,000 --> 00:13:01,000

Because it is now profitable, the AI companies want to build a lot more tokens, generate a lot more tokens,

因為現在有利可圖，AI 公司希望建立更多 token、生成更多 token，

00:13:01,000 --> 00:13:08,000

build more AI factories, which is the reason why compute demand here in Taiwan has skyrocketed.

建造更多 AI 工廠，這正是台灣算力需求急劇攀升的原因。

00:13:08,000 --> 00:13:15,000

It is precisely the reason why all of you are so busy and your businesses are doing so well.

這正是你們所有人都如此繁忙、業務蒸蒸日上的原因。

00:13:15,000 --> 00:13:18,000

In fact, that looks like some of your stock price.

事實上，那看起來就像你們某些人的股價走勢。

00:13:25,000 --> 00:13:30,000

The compute pattern has changed. Everything has changed.

運算模式已經改變。一切都改變了。

00:13:30,000 --> 00:13:34,000

So the first idea is that useful AI has arrived.

第一個核心概念是：真正有用的 AI 已經到來。

00:13:34,000 --> 00:13:40,000

AI is now a profit generator. AI is now a GDP generator.

AI 現在是利潤的引擎。AI 現在是 GDP 的引擎。

00:13:40,000 --> 00:13:44,000

Behind it is a whole new kind of computing pattern.

背後是一種全新的運算模式。

00:13:44,000 --> 00:13:47,000

Not just a large language model, but an agent.

不只是一個大型語言模型，而是一個 Agent（自主代理）。

00:13:47,000 --> 00:13:52,000

Today, almost everything we're going to talk about is going to be based on this.

今天，我們幾乎要談論的一切都將以此為基礎。

00:13:52,000 --> 00:13:55,000

So let me take a quick moment and show you what I'm talking about.

讓我花一點時間，向你們展示我所說的。

00:13:55,000 --> 00:14:01,000

Inside, this is an agent. It's an agent application.

內部，這就是一個 Agent。這是一個 Agent 應用程式。

00:14:01,000 --> 00:14:08,000

In the old days, this would be application. This would be code.

過去，這是應用程式，這是程式碼。

00:14:08,000 --> 00:14:11,000

And this would be operating system.

而這是作業系統。

00:14:11,000 --> 00:14:16,000

Application, code running inside an application, inside an operating system.

應用程式、在應用程式內運行的程式碼、在作業系統之內。

00:14:16,000 --> 00:14:25,000

Today, it is agent, which consists of a large language model, or many, sitting inside a harness.

今天，它是 Agent，由一個或多個大型語言模型，嵌入在一個 harness（框架）之中。

00:14:25,000 --> 00:14:31,000

And that harness helps it, orchestrates it to do productive work.

而那個 harness 幫助它、協調它來完成有生產力的工作。

00:14:31,000 --> 00:14:42,000

This is the input. When that input comes, it has to understand, observe, reason, act, use tools, use tools.

這是輸入。當輸入進來時，它必須理解、觀察、推理、行動、使用工具、使用工具。

00:14:42,000 --> 00:14:49,000

That tool could be a spreadsheet, web browser, a data processing engine, database engine, for example.

那個工具可以是試算表、網頁瀏覽器、資料處理引擎、資料庫引擎，等等。

00:14:49,000 --> 00:14:57,000

This is orchestrated. This harness orchestrates this routing of information.

這是被協調的。這個 harness 協調著資訊的路由傳遞。

00:14:57,000 --> 00:15:06,000

Every single time it touches either processing the context, understanding what is happening, reasoning about what to do,

每當它處理上下文、理解當下發生什麼、推理該做什麼，

00:15:06,000 --> 00:15:15,000

coming up with a plan that you can act, that it acts on, that orchestration path is orchestrated by some software.

制定可以執行的計畫並付諸行動時，那個協調路徑都由某個軟體來管理。

00:15:15,000 --> 00:15:19,000

And so this is fundamentally an agent.

所以這本質上就是一個 Agent。

00:15:19,000 --> 00:15:26,000

It deals with short-term memory, called working memory, long-term memory, just like we do, we have long-term memory.

它處理短期記憶（稱為工作記憶）和長期記憶，就像我們一樣，我們也有長期記憶。

00:15:26,000 --> 00:15:30,000

And so the memory management system is incredibly important.

因此，記憶管理系統極其重要。

00:15:30,000 --> 00:15:34,000

This entire system is called an agent.

整個系統稱為 Agent（自主代理）。

00:15:34,000 --> 00:15:45,000

The large language model is used to do the thinking, and the harness connects everything together, just like an operating system.

大型語言模型負責思考，harness 將一切連接在一起，就像作業系統一樣。

00:15:45,000 --> 00:15:49,000

And so this is the new computing model, and this is what an agent.

所以這就是新的運算模型，這就是 Agent 的本質。

00:15:49,000 --> 00:15:53,000

It could do incredible things. This is the big breakthrough.

它能做出令人難以置信的事情。這是重大突破。

00:15:53,000 --> 00:16:03,000

The simultaneous convergence of large language models that are now able to do a really good job thinking, reasoning, planning, using tools,

大型語言模型的同步匯聚——它們現在能夠很好地思考、推理、規劃、使用工具，

00:16:03,000 --> 00:16:12,000

and the fact that we have now these harnesses that manages memory, the orchestration, uses tools, we can now do amazing things.

再加上我們現在有了這些管理記憶、協調運作、使用工具的 harness，我們現在能夠做到驚人的事情。

00:16:12,000 --> 00:16:14,000

Let me give you some example.

讓我舉一些例子。

00:16:14,000 --> 00:16:18,000

This is a prompt. This is the prompt.

這是一個 prompt。這就是那個 prompt。

00:16:18,000 --> 00:16:24,000

This is the code that is generated, and this comes out.

這是生成的程式碼，這是輸出結果。

00:16:24,000 --> 00:16:27,000

This is the input.

這是輸入。

00:16:27,000 --> 00:16:32,000

This is the input, and that's the output.

這是輸入，那是輸出。

00:16:32,000 --> 00:16:39,000

What do you guys think? It's pretty amazing, right?

大家覺得怎麼樣？很驚人吧？

00:16:39,000 --> 00:16:43,000

We use cloud code here, but Codex does an incredible job as well.

我們這裡用的是 Claude Code，但 Codex 也做得非常出色。

00:16:43,000 --> 00:16:44,000

Here's another example.

這是另一個例子。

00:16:44,000 --> 00:16:45,000

This is the input.

這是輸入。

00:16:45,000 --> 00:17:01,000

Create a GIF, NVIDIA green dots on black scatter, form Taiwan 101 building, more of 2 GTC Taipei 2026, more of 2 NVIDIA iLogo, then scatter and repeat.

製作一個 GIF：黑色背景上散布 NVIDIA 綠色點，構成台北 101 大樓造型，變形為 GTC Taipei 2026 文字，再變形為 NVIDIA iLogo，然後散開並重複。

00:17:01,000 --> 00:17:04,000

Right? So you saw that. That was the prompt.

對吧？你們看到了。那就是那個 prompt。

00:17:04,000 --> 00:17:05,000

Here's the next one.

下一個。

00:17:05,000 --> 00:17:09,000

I lost my remote control battery clip. It looks like this.

我弄丟了遙控器的電池夾，它長這樣。

00:17:09,000 --> 00:17:13,000

Create a CAD file. It uses a tool. Create a CAD file.

製作一個 CAD 檔案。它使用了工具。製作一個 CAD 檔案。

00:17:13,000 --> 00:17:17,000

Ready for 3D printing to create a new one.

為了 3D 列印而準備，製作一個新的。

00:17:17,000 --> 00:17:19,000

Make sense?

明白了嗎？

00:17:19,000 --> 00:17:22,000

This is now the new computing pattern.

這就是全新的運算模式。

00:17:22,000 --> 00:17:28,000

Whereas we used to launch an application, click and type.

我們過去習慣啟動應用程式、點擊和輸入。

00:17:28,000 --> 00:17:42,000

We now replace that with explaining to the AI what we want, our intent, and the AI generates the code or uses tools and produce the necessary output.

現在我們用向 AI 說明我們想要什麼（我們的意圖）來取代這一切，AI 生成程式碼或使用工具，產生必要的輸出。

00:17:42,000 --> 00:17:46,000

This is how computers are going to work in the future.

這就是未來電腦的運作方式。

00:17:46,000 --> 00:17:48,000

This is agentic AI.

這就是 Agentic AI。

100

00:17:48,000 --> 00:17:53,000

For two years we've been building towards this, and now it has arrived.

兩年來我們一直在朝這個方向建構，現在它已經到來。

101

00:17:53,000 --> 00:17:56,000

Now, one of the big breakthroughs, of course, is tool use.

現在，其中一個重大突破當然是工具使用（tool use）。

102

00:17:56,000 --> 00:18:04,000

A lot of people have said, you know, Jensen, AI is coming, agentic AI is coming, therefore all of the software companies are going to go out of business.

很多人說：「Jensen，AI 要來了，Agentic AI 要來了，所以所有軟體公司都要倒閉了。」

103

00:18:04,000 --> 00:18:07,000

I said it's exactly the opposite.

我說恰恰相反。

104

00:18:07,000 --> 00:18:14,000

Because there are going to be so many agents, the world is no longer limited by the number of people.

因為將會有無數的 Agent，世界不再受人數所限制。

105

00:18:14,000 --> 00:18:20,000

Therefore, those agents are going to use more tools than ever.

因此，那些 Agent 將比以往使用更多工具。

106

00:18:20,000 --> 00:18:24,000

This is actually an incredible time to be a software company.

這實際上是身為軟體公司的最佳時機。

107

00:18:24,000 --> 00:18:30,000

But the software has to be presented to the agent in a way that the agent can use it.

但軟體必須以 Agent 能夠使用的方式呈現給 Agent。

108

00:18:30,000 --> 00:18:32,000

This is a big breakthrough.

這是一個重大突破。

109

00:18:32,000 --> 00:18:40,000

And in fact, what we have done, as you know, what NVIDIA's treasure is, is all of our CUDA libraries.

事實上，我們所做的，你們都知道，NVIDIA 的寶藏，就是我們所有的 CUDA 函式庫。

110

00:18:40,000 --> 00:18:42,000

I call them CUDA-X libraries.

我稱它們為 CUDA-X 函式庫。

111

00:18:42,000 --> 00:18:44,000

This is NVIDIA's treasure.

這就是 NVIDIA 的寶藏。

112

00:18:44,000 --> 00:18:55,000

Today, we're able to now present these CUDA-X libraries to agents who can use it much more effectively than even humans.

今天，我們能夠將這些 CUDA-X 函式庫呈現給 Agent，Agent 能夠比人類更有效率地使用它們。

113

00:18:55,000 --> 00:18:58,000

And so this is a wonderful time for CUDA-X libraries.

所以這是 CUDA-X 函式庫的美好時代。

114

00:18:58,000 --> 00:19:00,000

Let's take a look.

讓我們來看看。

115

00:19:00,000 --> 00:19:08,000

Twenty years ago, we built CUDA, a single architecture for accelerated computing.

二十年前，我們建構了 CUDA，一個用於加速運算的統一架構。

116

00:19:08,000 --> 00:19:11,000

We reinvented computing.

我們重新發明了運算。

117

00:19:11,000 --> 00:19:18,000

A thousand CUDA-X libraries help developers make breakthroughs in every field of science and engineering.

一千個 CUDA-X 函式庫幫助開發者在科學與工程的每個領域取得突破。

118

00:19:18,000 --> 00:19:22,000

CUDA-X libraries are tools for agents.

CUDA-X 函式庫是 Agent 的工具。

119

00:19:22,000 --> 00:19:26,000

CULITO for computational lithography.

CULITO 用於計算微影技術。

120

00:19:26,000 --> 00:19:31,000

CUOBT for decision optimization.

CUOBT 用於決策最佳化。

121

00:19:31,000 --> 00:19:35,000

CUDSS for direct sparse solvers.

CUDSS 用於直接稀疏求解器。

122

00:19:35,000 --> 00:19:41,000

AIQ for deep research across structured and unstructured documents.

AIQ 用於跨結構化與非結構化文件的深度研究。

123

00:19:41,000 --> 00:19:45,000

ARIEL for AI-RAN.

ARIEL 用於 AI-RAN（AI 無線接取網路）。

124

00:19:45,000 --> 00:19:49,000

WARP for differentiable physics.

WARP 用於可微分物理模擬。

125

00:19:49,000 --> 00:19:52,000

PARABRIX for genomics.

PARABRIX 用於基因體學。

126

00:19:52,000 --> 00:19:55,000

At their foundation are algorithms.

它們的基礎是演算法。

127

00:19:55,000 --> 00:19:58,000

And they are beautiful.

而且它們非常美麗。

128

00:22:48,000 --> 00:22:53,000

A round of applause for math.

為數學鼓掌。

129

00:22:53,000 --> 00:22:56,000

Math is beautiful.

數學是美麗的。

130

00:22:56,000 --> 00:23:04,000

The computing pattern, the computing pattern of software is going to change.

運算模式，軟體的運算模式即將改變。

131

00:23:04,000 --> 00:23:07,000

In fact, let's come back to this.

事實上，讓我們回到這個話題。

132

00:23:07,000 --> 00:23:09,000

This is the agent.

這就是 Agent。

133

00:23:09,000 --> 00:23:18,000

This is the ultimate disaggregated and distributed computing, computing model.

這是終極的分解式與分散式運算模型。

134

00:23:18,000 --> 00:23:24,000

So many different computers are going to be activated in order to process this agent.

為了處理這個 Agent，將會啟動非常多不同的電腦。

135

00:23:24,000 --> 00:23:35,000

The agent consists of model, harness, tools and skills, and a runtime.

Agent 由模型、harness、工具與技能，以及 runtime 所組成。

136

00:23:35,000 --> 00:23:40,000

All of that is running at different places in a data center.

所有這些都在資料中心的不同位置運行。

137

00:23:40,000 --> 00:23:47,000

You can think of the model as the brain, the harness as the body,

你可以把模型想成大腦，把 harness 想成身體，

138

00:23:47,000 --> 00:23:54,000

the tools that it uses, working in a runtime, think of it as a workshop.

它使用的工具在 runtime 中運作，把它想成一個工作坊。

139

00:23:54,000 --> 00:24:00,000

So this is a person, a worker working with tools in a workshop.

所以這就像一個人、一個工人，在工作坊裡使用工具工作。

140

00:24:00,000 --> 00:24:04,000

Of course, this is being done at extraordinarily large scales.

當然，這是在極其龐大的規模下進行的。

141

00:24:04,000 --> 00:24:09,000

And each one of those steps are running in a different part of the computer.

每一個步驟都在電腦的不同部分運行。

142

00:24:09,000 --> 00:24:15,000

And you could see the large language model is thinking, context processing,

你可以看到，大型語言模型正在思考、處理上下文，

143

00:24:15,000 --> 00:24:21,000

observing, understanding the environment, reasoning, coming up with a plan,

觀察、理解環境、推理、制定計畫，

144

00:24:21,000 --> 00:24:23,000

and acting on the plan.

並執行計畫。

145

00:24:23,000 --> 00:24:31,000

Every single time that happens, an entire rack of Grace Blackwell NVLink 72 is activated.

每一次這個過程發生，都會啟動一整個機架的 Grace Blackwell NVLink 72。

146

00:24:31,000 --> 00:24:34,000

And this is thinking with the large language model.

這就是用大型語言模型進行思考。

147

00:24:34,000 --> 00:24:38,000

Whenever it uses a tool, a CPU is used.

每當它使用工具時，就會用到 CPU。

148

00:24:38,000 --> 00:24:44,000

That tool could be a C compiler, it could be Python, it could be JavaScript,

那個工具可以是 C 編譯器、Python 或 JavaScript，

149

00:24:44,000 --> 00:24:47,000

or it could be accelerated computing.

或者是加速運算。

150

00:24:47,000 --> 00:24:52,000

Today's agents are relatively simple users of tools.

今日的 Agent 在使用工具方面還相對簡單。

151

00:24:52,000 --> 00:24:56,000

Tomorrow, they're going to be very sophisticated users of tools,

明日，它們將成為非常精通工具的使用者，

152

00:24:56,000 --> 00:25:02,000

which is the reason why the CUDA-X libraries that I showed you are going to be incredibly popular with agents.

這正是我向你們展示的 CUDA-X 函式庫將在 Agent 中大受歡迎的原因。

153

00:25:02,000 --> 00:25:05,000

They solve some of the most important problems the world knows.

它們解決了世界上一些最重要的問題。

154

00:25:05,000 --> 00:25:15,000

And all of our CUDA-X libraries are now going to come with skills that the AI could learn how to use.

我們所有的 CUDA-X 函式庫現在都將附帶技能（skills），讓 AI 學會如何使用它們。

155

00:25:15,000 --> 00:25:24,000

So the CUDA-X library, some skills, basically a manual, the AI reads it and goes, "Aha, that's how you use it."

所以 CUDA-X 函式庫加上一些技能，基本上就是一本使用手冊，AI 讀完就說：「啊哈，原來是這樣用的。」

156

00:25:24,000 --> 00:25:28,000

The ability to use these libraries by agents are going to be incredible.

Agent 使用這些函式庫的能力將會是驚人的。

157

00:25:28,000 --> 00:25:34,000

And so the tools run on CPUs and GPUs and large language models.

工具在 CPU、GPU 和大型語言模型上運行。

158

00:25:34,000 --> 00:25:43,000

The security harness runs on CPUs and a security processor called a DPU, NVIDIA's Bluefield.

安全 harness 在 CPU 和稱為 DPU（NVIDIA 的 Bluefield）的安全處理器上運行。

159

00:25:43,000 --> 00:25:46,000

The orchestration of all this runs on a CPU.

所有這些的協調都在 CPU 上運行。

160

00:25:46,000 --> 00:25:51,000

This is the entire harness and the CPUs orchestrating all of the work.

這就是完整的 harness 和 CPU 協調所有工作的方式。

161

00:25:51,000 --> 00:25:54,000

One of the hardest parts is memory.

其中最困難的部分之一是記憶體（memory）。

162

00:25:54,000 --> 00:25:58,000

You could just imagine the working memory is called KV caching.

你可以想像，工作記憶體稱為 KV caching（鍵值快取）。

163

00:25:58,000 --> 00:26:04,000

What to remember? Compaction, not just compression, but how to retrieve.

要記住什麼？壓縮——不只是資料壓縮，還包括如何檢索。

164

00:26:04,000 --> 00:26:08,000

Do you retrieve structured data? Do you retrieve unstructured data?

你是檢索結構化資料？還是非結構化資料？

165

00:26:08,000 --> 00:26:14,000

What is the ontology, the relationship of all of these different data to itself?

這些不同資料彼此之間的本體論、關聯性是什麼？

166

00:26:14,000 --> 00:26:17,000

That entire processing is incredibly complicated.

整個處理過程非常複雜。

167

00:26:17,000 --> 00:26:26,000

The memory system, the memory system of AIs is going to cause the storage system to be completely revolutionized.

AI 的記憶系統將導致存儲系統被徹底革新。

168

00:26:26,000 --> 00:26:36,000

As you can see, every aspect of this computing model, this computing pattern, this new application called an agent,

如你所見，這個運算模型的每個層面，這個運算模式，這個稱為 Agent 的新應用，

169

00:26:36,000 --> 00:26:40,000

is fundamentally different than the way that applications used to run.

與過去應用程式的運行方式根本不同。

170

00:26:40,000 --> 00:26:46,000

A whole bunch of software sitting inside a binary, sitting inside an operating system.

過去是大量軟體嵌在二進制檔案中，在作業系統裡運行。

171

00:26:46,000 --> 00:27:00,000

This is the reason this disaggregated, this distributed, this heterogeneous computing problem is precisely the reason we built our next generation.

這種分解式、分散式、異質運算的需求，正是我們打造下一代產品的原因。

172

00:27:00,000 --> 00:27:02,000

Vera Rubin.

Vera Rubin。

173

00:27:02,000 --> 00:27:05,000

Vera Rubin is not one chip.

Vera Rubin 不是一顆晶片。

174

00:27:05,000 --> 00:27:08,000

Vera Rubin is not a GPU only.

Vera Rubin 不只是一顆 GPU。

175

00:27:08,000 --> 00:27:14,000

It starts with the GPU, but Vera Rubin is incredible.

它從 GPU 開始，但 Vera Rubin 遠不止於此。

176

00:27:14,000 --> 00:27:20,000

This entire thing is Vera Rubin from end to end.

整個這一切都是 Vera Rubin，從頭到尾。

177

00:27:20,000 --> 00:27:25,000

It has GPUs, Vera Rubin, NVLink 72.

它有 GPU、Vera Rubin、NVLink 72。

178

00:27:25,000 --> 00:27:29,000

It is orchestrated by Vera CPUs that I'm going to tell you more about.

它由 Vera CPU 協調，我稍後會詳細介紹。

179

00:27:29,000 --> 00:27:32,000

The storage systems, revolutionary.

存儲系統，革命性的。

180

00:27:32,000 --> 00:27:48,000

Vera, along with CX-9, our software stack called DoCA, the security processor that's inside so that everything is encrypted at rest, in motion, as well as in use.

Vera 加上 CX-9，我們稱為 DoCA 的軟體堆疊，內建安全處理器，使所有內容在靜止、傳輸和使用中都加密。

181

00:27:48,000 --> 00:27:53,000

Everything across this is secure because the AI model is so precious.

整個系統的安全性極高，因為 AI 模型實在太珍貴了。

182

00:27:53,000 --> 00:27:59,000

This is the reason why this entire system obeys confidential computing.

這就是整個系統遵循機密運算（confidential computing）的原因。

183

00:27:59,000 --> 00:28:04,000

Each one of these systems would be a complete revolution in itself.

其中每一個系統本身就是一場徹底的革命。

184

00:28:04,000 --> 00:28:10,000

Vera Rubin is the most ambitious endeavor in the history of our company.

Vera Rubin 是我們公司歷史上最雄心勃勃的計畫。

185

00:28:10,000 --> 00:28:18,000

The whole company worked on Vera Rubin across all 40,000 engineers, not to mention all of you.

全公司 4 萬名工程師都投入了 Vera Rubin 的開發，更不用說你們所有人了。

186

00:28:18,000 --> 00:28:22,000

All of you participated in the creation of this entire system.

你們所有人都參與了整個系統的創建。

187

00:28:22,000 --> 00:28:26,000

Vera Rubin is really a miracle, and it's not just one chip.

Vera Rubin 真的是一個奇蹟，而且它不只是一顆晶片。

188

00:28:26,000 --> 00:28:28,000

It is so many.

它涵蓋了太多東西。

189

00:28:28,000 --> 00:28:31,000

Well, it's even beyond that.

好，它甚至超越了這些。

190

00:28:31,000 --> 00:28:34,000

A long time ago, NVIDIA used to be a GPU company.

很久以前，NVIDIA 曾經是一家 GPU 公司。

191

00:28:34,000 --> 00:28:40,000

But over the years, we've evolved to become a systems company.

但多年來，我們已進化成為一家系統公司。

192

00:28:40,000 --> 00:28:47,000

You're looking here now for the most complex systems, most complex and ground-up system ever designed.

你們現在看到的是有史以來設計過最複雜、最從零開始打造的系統。

193

00:28:47,000 --> 00:28:54,000

But ultimately, our customers, our partners, don't want to buy computers.

但歸根究柢，我們的客戶、我們的合作夥伴不想買電腦。

194

00:28:54,000 --> 00:29:02,000

They want to build AI factories, which is the reason why NVIDIA has really started to transform ourselves yet again.

他們想建 AI 工廠，這就是 NVIDIA 為什麼再次開始轉型的原因。

195

00:29:02,000 --> 00:29:08,000

You can see so much of our technology is now at the entire infrastructure scale.

你可以看到，我們的大量技術現在已達到整個基礎架構的規模。

196

00:29:08,000 --> 00:29:11,000

Our partners are at infrastructure scale.

我們的合作夥伴也在基礎架構規模上運作。

197

00:29:11,000 --> 00:29:16,000

Power generators, cooling systems, the grid providers.

發電設備、冷卻系統、電網供應商。

198

00:29:16,000 --> 00:29:21,000

So many industrial companies are now part of our ecosystem

現在有如此多的工業公司成為我們生態系統的一部分，

199

00:29:21,000 --> 00:29:27,000

because ultimately, we're trying to build an entire stack, just like GPUs,

因為歸根究柢，我們試圖打造一整個完整的技術堆疊，就像 GPU 一樣，

200

00:29:27,000 --> 00:29:31,000

just like when we were building Grace Blackwell NVLink 72,

就像我們建造 Grace Blackwell NVLink 72 時一樣，

201

00:29:31,000 --> 00:29:40,000

just like now we are building a full stack system so that our customers could build amazing AI infrastructure.

就像我們現在正在建造完整的全堆疊系統，讓客戶能夠建造驚人的 AI 基礎架構。

202

00:29:40,000 --> 00:29:42,000

Let's take a look.

讓我們來看看。

203

00:29:42,000 --> 00:29:49,000

The world is racing to build AI factories, the largest infrastructure build out in human history.

全世界正在競相建造 AI 工廠，這是人類歷史上規模最大的基礎架構建設。

204

00:29:49,000 --> 00:29:53,000

AI factories are incredibly complex, every layer.

AI 工廠極其複雜，每一層都是。

205

00:29:53,000 --> 00:30:03,000

Chip, rack, network, power, cooling, grid must be designed together from end to end because compute is revenues.

晶片、機架、網路、電力、冷卻、電網必須從頭到尾整合設計，因為算力就是營收。

206

00:30:03,000 --> 00:30:14,000

NVIDIA DSX is the blueprint, a reference design for building and operating AI factories at maximum efficiency and profitability.

NVIDIA DSX 是藍圖，是以最高效率和最大獲利能力建造和運營 AI 工廠的參考設計。

207

00:30:14,000 --> 00:30:17,000

It starts with DSX Sim.

從 DSX Sim 開始。

208

00:30:17,000 --> 00:30:26,000

With the DSX Sim Omniverse Blueprint, partners design and validate an NVIDIA VeraRubin AI factory before a single rack lands.

透過 DSX Sim Omniverse Blueprint，合作夥伴在第一個機架安裝前，就能設計和驗證 NVIDIA VeraRubin AI 工廠。

209

00:30:26,000 --> 00:30:42,000

They plan the layout, simulate the power and cooling, design the network, validate every integration, test every change in the digital twin.

他們規劃佈局、模擬電力和冷卻、設計網路、驗證每個整合環節，並在數位孿生中測試每一項變更。

210

00:30:42,000 --> 00:30:44,000

The factory powers on.

工廠啟動。

211

00:30:44,000 --> 00:30:51,000

DSX OS takes over and provisions, operates, monitors and remediates the infrastructure,

DSX OS 接管，負責部署、運營、監控和修復基礎架構，

212

00:30:51,000 --> 00:30:59,000

turning the installed systems into trusted, multi-tenant, resilient, AI-ready capacity.

將安裝的系統轉化為可信任、多租戶、具韌性、隨時可用的 AI 算力。

213

00:30:59,000 --> 00:31:04,000

Today's AI factories over-provision power by up to 40%.

今日的 AI 工廠超額預留最多 40% 的電力。

214

00:31:04,000 --> 00:31:11,000

DSX Max LPS lets operators safely deploy more GPUs inside the same power budget,

DSX Max LPS 讓運營商在相同的電力預算內安全部署更多 GPU，

215

00:31:11,000 --> 00:31:14,000

adding billions in annual revenue.

每年增加數十億美元的收入。

216

00:31:14,000 --> 00:31:22,000

Breakthrough hot liquid cooling at 45 degrees Celsius uses less water and energy.

突破性的 45 攝氏度熱液體冷卻技術，使用更少的水和能源。

217

00:31:22,000 --> 00:31:25,000

More power going to revenue-generating compute.

更多電力用於創造收益的運算。

218

00:31:25,000 --> 00:31:27,000

Incredible.

令人難以置信。

219

00:31:27,000 --> 00:31:36,000

Dynamic power allocation steers power from rack to rack, recovering stranded watts, sending them where work is happening.

動態功率分配將電力從機架調配到機架，回收閒置瓦特，將它們送往有工作負載的地方。

220

00:31:36,000 --> 00:31:42,000

In-rack power smoothing flattens peak current spikes and power surges.

機架內功率平滑化，消除峰值電流尖峰和電力突波。

221

00:31:42,000 --> 00:31:49,000

Throughout the factory, teams of AI agents work with DSX Max LPS,

整個工廠中，AI Agent 團隊與 DSX Max LPS 協同工作，

222

00:31:49,000 --> 00:31:54,000

continuously coordinating to balance cooling and power to meet workload demand.

持續協調，平衡冷卻與電力，以滿足工作負載需求。

223

00:31:54,000 --> 00:32:01,000

DSX AI factories are flexible energy assets that operate cooperatively with the grid.

DSX AI 工廠是靈活的能源資產，與電網協同運作。

224

00:32:01,000 --> 00:32:08,000

DSX Flex reads real-time grid signals and dynamically adjusts factory power when the grid needs relief.

DSX Flex 讀取即時電網訊號，在電網需要減壓時動態調整工廠用電量。

225

00:32:08,000 --> 00:32:16,000

100 gigawatts of AI factories will come online before the end of the decade.

在十年結束前，將有 100 GW 的 AI 工廠上線運作。

226

00:32:16,000 --> 00:32:24,000

NVIDIA DSX AI factories run at highest efficiency, produce the lowest cost tokens and make the grid stronger.

NVIDIA DSX AI 工廠以最高效率運行，產出成本最低的 token，並讓電網更強壯。

227

00:32:32,000 --> 00:32:37,000

I've shown you ecosystem slides of the past,

我過去曾給你們看過生態系統的投影片，

228

00:32:37,000 --> 00:32:42,000

where NVIDIA's computing layers and software and software and computing stacks

展示 NVIDIA 的運算層和軟體、軟體和運算堆疊

229

00:32:42,000 --> 00:32:49,000

are integrated into other people's platforms, third-party platforms and libraries that serves end markets.

被整合進其他人的平台、第三方平台和服務終端市場的函式庫。

230

00:32:49,000 --> 00:32:52,000

That was a computing ecosystem.

那是一個運算生態系統。

231

00:32:52,000 --> 00:32:55,000

This is an AI factory ecosystem.

這是一個 AI 工廠生態系統。

232

00:32:55,000 --> 00:32:58,000

This is way downstream of all of you.

這遠在你們所有人的下游。

233

00:32:58,000 --> 00:33:04,000

Upstream of me is all of you, and downstream of us is this ecosystem.

我的上游是你們所有人，而我們的下游是這個生態系統。

234

00:33:04,000 --> 00:33:10,000

Because NVIDIA ultimately is not just building a GPU, not just building a system,

因為 NVIDIA 最終不只是在建造 GPU，不只是在建造系統，

235

00:33:10,000 --> 00:33:17,000

we're helping customers build these AI factories, these AI infrastructure that is so immensely complex.

我們是在幫助客戶建造這些 AI 工廠，這些極其複雜的 AI 基礎架構。

236

00:33:17,000 --> 00:33:25,000

Each one of these at one gigawatt level started at 30, 20, 30 billion dollars.

每個 1 GW 規模的 AI 工廠造價從 200 億、300 億美元起跳。

237

00:33:25,000 --> 00:33:33,000

It is at 50, 60 billion dollars, and soon it will be 80, 100 billion dollars per gigawatt.

現在是 500 億、600 億美元，很快就會達到每 GW 800 億到 1,000 億美元。

238

00:33:33,000 --> 00:33:38,000

100 billion dollars into an AI factory.

1,000 億美元投入一座 AI 工廠。

239

00:33:38,000 --> 00:33:42,000

It must work the first time, and it must work right away.

它必須一次就成功，而且必須立刻投入運作。

240

00:33:42,000 --> 00:33:44,000

The cost of capital is incredible.

資本成本是令人難以置信的。

241

00:33:44,000 --> 00:33:47,000

The complexity is incredible.

複雜度也是令人難以置信的。

242

00:33:47,000 --> 00:33:52,000

So as you see, we used to design a chip inside a computer.

所以如你所見，我們過去在電腦裡設計晶片。

243

00:33:52,000 --> 00:33:56,000

And then we simulated a system inside a computer.

然後我們在電腦裡模擬系統。

244

00:33:56,000 --> 00:34:03,000

Today, you saw just now, everything was built in Omniverse.

今天，你們剛才看到了，一切都在 Omniverse 中建造。

245

00:34:03,000 --> 00:34:06,000

I've been working with Omniverse with all of you for a long time.

我和你們一起使用 Omniverse 已經很長時間了。

246

00:34:06,000 --> 00:34:08,000

This was the dream come true.

這是夢想成真。

247

00:34:08,000 --> 00:34:14,000

So that we can build these gigantic systems as large as the world wants to build.

這樣我們就能建造這些巨型系統，規模隨世界需求而定。

248

00:34:14,000 --> 00:34:20,000

Inside a digital framework, inside a digital simulator, in a digital world.

在數位框架內、在數位模擬器內、在數位世界中。

249

00:34:20,000 --> 00:34:25,000

Long before we built the first great ground and put our money to work.

遠在我們打下第一個地基、投入資金之前。

250

00:34:25,000 --> 00:34:29,000

So this is our ecosystem. We call it DSX.

所以這就是我們的生態系統。我們稱之為 DSX。

251

00:34:29,000 --> 00:34:37,000

RTX is for our GPU, DGX is for our systems, and now DSX, basically infrastructure.

RTX 是我們的 GPU 品牌，DGX 是我們的系統品牌，現在 DSX 基本上是基礎架構品牌。

252

00:34:37,000 --> 00:34:43,000

Because of the work that we do here, across this entire stack, including our systems and software,

正因為我們在整個技術堆疊所做的工作，包括我們的系統和軟體，

253

00:34:43,000 --> 00:34:50,000

it's the reason why we can work with small companies and enable them to be world-class AI clouds.

這就是我們能與小公司合作，讓它們成為世界級 AI 雲端業者的原因。

254

00:34:50,000 --> 00:34:55,000

Every one of these I'm about to show you are small companies just recently.

我即將展示的每一家，最近都還只是小公司。

255

00:34:55,000 --> 00:35:01,000

And now CoreWeave is worth 50, 60, 70 billion dollars and growing incredibly fast.

而現在 CoreWeave 的估值已達 500 億、600 億、700 億美元，且成長極其迅速。

256

00:35:01,000 --> 00:35:06,000

Recently, we worked with Nebius, and again, they're growing incredibly fast.

最近，我們與 Nebius 合作，同樣，他們的成長也是極其迅速。

257

00:35:06,000 --> 00:35:10,000

Each one of these clouds have incredible customers.

這些雲端業者都有令人驚嘆的客戶。

258

00:35:10,000 --> 00:35:16,000

Cursor, the software coding company, Black Mountain Labs, Image Generation, World Labs,

Cursor（軟體編程公司）、Black Mountain Labs（圖像生成）、World Labs，

259

00:35:16,000 --> 00:35:23,000

World Foundation Model, Revolut, the leading financial services AI company, and Shopify.

World Foundation Model、Revolut（領先的金融服務 AI 公司），以及 Shopify。

260

00:35:23,000 --> 00:35:30,000

Here's another one. This is NScale, and their customers are British Telecom, Google.

這是另一家。這是 NScale，他們的客戶包括英國電信（British Telecom）、Google。

261

00:35:30,000 --> 00:35:38,000

Google is using one of our AI clouds. Thinking Machines, a Frontier Labs company, super exciting.

Google 正在使用我們的一個 AI 雲端。Thinking Machines，一家 Frontier Labs 的公司，超令人興奮。

262

00:35:38,000 --> 00:35:46,000

Here's Naver Cloud in Korea, Bank of Korea, Hyundai, so many incredible companies.

這是韓國的 Naver Cloud，還有韓國銀行、現代汽車，這麼多令人驚嘆的公司。

263

00:35:46,000 --> 00:35:51,000

Here's one in India, Yoda, incredible companies.

這是印度的一家，Yotta，令人驚嘆的公司。

264

00:35:51,000 --> 00:35:59,000

Here's one based in Singapore, building in Australia, Together AI, AI Singapore.

這是一家總部在新加坡、在澳洲建設的，Together AI、AI Singapore。

265

00:35:59,000 --> 00:36:01,000

This is one in Indonesia.

這是印尼的一家。

266

00:36:01,000 --> 00:36:08,000

Each one of these companies are serving regional as well as global customers.

這些公司每一家都在服務區域性和全球性客戶。

267

00:36:08,000 --> 00:36:13,000

AI is going to run everywhere. Every company will be powered by it.

AI 將無處不在。每家公司都將由它驅動。

268

00:36:13,000 --> 00:36:17,000

Every region will build it.

每個地區都將建造它。

269

00:36:17,000 --> 00:36:23,000

Indosat here in Indonesia. Here in Taiwan, GMI.

印尼的 Indosat。台灣的 GMI。

270

00:36:23,000 --> 00:36:28,000

Here in Taiwan, GMI. It's okay to clap.

台灣的 GMI。可以鼓掌的。

271

00:36:33,000 --> 00:36:41,000

So, incredible, incredible companies, incredible opportunity, but all of them need several things.

所以，令人驚嘆的公司、令人驚嘆的機遇，但它們都需要幾樣東西。

272

00:36:41,000 --> 00:36:45,000

Of course, they need the computing stack. This entire stack underneath.

當然，它們需要運算堆疊。這整個底層堆疊。

273

00:36:45,000 --> 00:36:50,000

This is what made NVIDIA famous. All of our hardware and software and libraries,

這就是讓 NVIDIA 聞名的東西。我們所有的硬體、軟體和函式庫，

274

00:36:50,000 --> 00:36:55,000

our connection into the world's ecosystem of third-party developers,

我們與全球第三方開發者生態系統的連結，

275

00:36:55,000 --> 00:37:00,000

makes it possible for anyone to stand up an AI cloud.

讓任何人都能建立 AI 雲端服務。

276

00:37:00,000 --> 00:37:04,000

However, the AI cloud is so complex now.

然而，AI 雲端現在已經非常複雜。

277

00:37:04,000 --> 00:37:08,000

This is the software version. This is the computer science version.

這是軟體版本。這是電腦科學版本。

278

00:37:08,000 --> 00:37:14,000

The money version, the asset version, is what I showed you earlier.

金融版本、資產版本，是我之前展示給你們看的。

279

00:37:14,000 --> 00:37:20,000

It's a giant factory. Having this ability alone is not enough,

那是一個巨型工廠。光有這種能力是不夠的，

280

00:37:20,000 --> 00:37:24,000

which is the reason why NVIDIA has become an AI infrastructure company.

這就是 NVIDIA 為什麼已成為一家 AI 基礎架構公司的原因。

281

00:37:24,000 --> 00:37:33,000

Now, doing this well and becoming incredibly good at helping customers build AI factories

現在，把這件事做好，成為幫助客戶建造 AI 工廠的頂尖專家，

282

00:37:33,000 --> 00:37:36,000

and deploying AI factories is incredibly important.

以及部署 AI 工廠，是極其重要的。

283

00:37:36,000 --> 00:37:41,000

And the reason for that is this. Compute is revenue now.

原因就是這個。算力現在就是收益。

284

00:37:41,000 --> 00:37:47,000

Compute is profit. The absence of revenues and profit is loss.

算力就是利潤。沒有收益和利潤就是損失。

285

00:37:47,000 --> 00:37:58,000

And so, it's really important to realize that this is when this is an example of an AI infrastructure coming online.

所以，真的很重要的是要意識到，這是一個 AI 基礎架構上線的例子。

286

00:37:58,000 --> 00:38:03,000

It could be coming online quickly. It could take a while.

它可能快速上線，也可能需要一段時間。

287

00:38:03,000 --> 00:38:10,000

Its throughput could be high. It could be low. Its resilience and reliability could be good or bad.

其吞吐量可高可低。其韌性和可靠性可好可壞。

288

00:38:10,000 --> 00:38:15,000

And its lifetime of usefulness could be long or short.

其有用壽命可長可短。

289

00:38:15,000 --> 00:38:23,000

Because this represents 50, 60, going to 100 billion dollars.

因為這代表 500 億、600 億、進而達到 1,000 億美元的投資。

290

00:38:23,000 --> 00:38:29,000

This curve matters greatly, which is the reason why NVIDIA is such a great partner.

這條曲線至關重要，這就是為什麼 NVIDIA 是如此出色的合作夥伴。

291

00:38:29,000 --> 00:38:37,000

Working with us because of our fully integrated capability, we didn't just come up with a PowerPoint slide.

與我們合作，因為我們完全整合的能力——我們不只是做了一份 PowerPoint 投影片。

292

00:38:37,000 --> 00:38:41,000

We created the entire infrastructure. We connected everything together.

我們建造了整個基礎架構。我們將一切連接在一起。

293

00:38:41,000 --> 00:38:48,000

We built out billions and billions of it ourselves to make sure that everything works well.

我們自己投入了數十億美元來確保一切運作良好。

294

00:38:48,000 --> 00:38:58,000

As a result of that, our time to first token, our time to first token, our time to first inference,

因此，我們的首個 token 時間、首個 token 時間、首次推理時間，

295

00:38:58,000 --> 00:39:04,000

our time to training turned on is much faster.

我們的訓練啟動時間要快得多。

296

00:39:04,000 --> 00:39:14,000

Second, because our throughput per watt, our tokens per watt is utterly world class.

第二，因為我們的每瓦吞吐量、每瓦 token 數絕對是世界級的。

297

00:39:14,000 --> 00:39:19,000

And the reason for that is because we integrate everything. We design everything to ground up.

原因是我們整合了一切。我們從頭開始設計一切。

298

00:39:19,000 --> 00:39:23,000

We simulate the entire system and we use extreme co-design.

我們模擬整個系統，並採用極致協同設計（extreme co-design）。

299

00:39:23,000 --> 00:39:31,000

Just like I showed you just now with the Vera Rubin Rack, everything was designed in order to deliver on this incredible throughput.

就像我剛才展示的 Vera Rubin 機架，一切都是為了實現這種令人難以置信的吞吐量而設計的。

300

00:39:31,000 --> 00:39:40,000

If your data center, if your factory has one gigawatt, it will not have more.

如果你的資料中心、你的工廠有一 GW 的電力，它就不會有更多了。

301

00:39:40,000 --> 00:39:46,000

One gigawatt means one gigawatt. That's all the power generation you could do.

一 GW 就是一 GW。那是你所能有的全部電力。

302

00:39:46,000 --> 00:39:56,000

If you have one gigawatt of power, then throughput per watt is revenues because every token is profitable.

如果你有一 GW 的電力，那麼每瓦吞吐量就是收益，因為每個 token 都有利潤。

303

00:39:56,000 --> 00:40:06,000

Every token is revenues. This is the future. Compute is revenues. Performance per watt is your revenues.

每個 token 都是收益。這就是未來。算力就是收益。每瓦性能就是你的收益。

304

00:40:06,000 --> 00:40:15,000

Choosing the wrong architecture just because the chips are cheaper doesn't translate, doesn't make sense.

僅僅因為晶片便宜就選擇錯誤的架構，這說不通，毫無意義。

305

00:40:15,000 --> 00:40:22,000

You need to make sure that your revenues per watt, the more you buy, the more you make.

你需要確保每瓦收益——買得越多，賺得越多。

306

00:40:22,000 --> 00:40:30,000

And so tokens per watt. And then lastly, very, oh second, third is reliability.

所以是每瓦 token 數。然後最後——不對，第二、第三是可靠性。

307

00:40:30,000 --> 00:40:37,000

If you ever get a chance to see these data centers, there are so many moving parts, millions of cables.

如果你有機會看到這些資料中心，有如此多的運動部件，數百萬根電纜。

308

00:40:37,000 --> 00:40:45,000

The ability for all of those computers to work harmoniously, reliability is extremely low.

要讓所有這些電腦協調工作，可靠性極難保證。

309

00:40:45,000 --> 00:40:51,000

It is just extremely difficult. We have now been operating very large scale for a very long time.

這真的極其困難。我們已在極大規模下運作了很長時間。

310

00:40:51,000 --> 00:40:59,000

That experience matters. That difference, mean time between interrupts, extremely important.

那個經驗非常重要。這個差異——中斷平均間隔時間（MTBI）——極其重要。

311

00:40:59,000 --> 00:41:10,000

And then lastly, this is very hard. The lifetime of these systems, the software is changing all the time.

最後，這非常困難。這些系統的壽命——軟體一直在持續變化。

312

00:41:10,000 --> 00:41:17,000

Four years ago, which is in the time of Hopper, AI has completely changed.

四年前，也就是 Hopper 時代，AI 已經完全改變了。

313

00:41:17,000 --> 00:41:24,000

Six years ago, this is the time frame of Ampere, AI has completely changed.

六年前，也就是 Ampere 的時代，AI 已經完全改變了。

314

00:41:24,000 --> 00:41:30,000

We started out talking about CNNs. Here we are. Then we talked about transformers.

我們一開始談的是 CNN（卷積神經網路）。看看現在。然後我們談到了 Transformer。

315

00:41:30,000 --> 00:41:35,000

And then we talked about a mixture of experts. Now we're talking about agentic systems.

然後我們談到了混合專家模型（Mixture of Experts）。現在我們在談 Agentic 系統。

316

00:41:35,000 --> 00:41:44,000

Every single generation, every single few months, the software industry is coming up with new technology.

每一個世代，每隔幾個月，軟體行業就會推出新技術。

317

00:41:44,000 --> 00:41:54,000

If your architecture is not flexible, if your ecosystem is not rich, then this curve cannot be long.

如果你的架構不夠靈活，如果你的生態系統不夠豐富，那這條曲線就不能持久。

318

00:41:54,000 --> 00:42:00,000

You cannot predict how long your system can last. I can.

你無法預測你的系統能撐多久。但我可以。

319

00:42:00,000 --> 00:42:06,000

NVIDIA systems is all over the world. Software developers start with NVIDIA CUDA.

NVIDIA 系統遍布全世界。軟體開發者從 NVIDIA CUDA 開始。

320

00:42:06,000 --> 00:42:15,000

And by definition, therefore, the life, the ecosystem, the useful asset is going to be much longer.

因此，從定義上來說，其壽命、生態系統、有效資產的使用期限將更長。

321

00:42:15,000 --> 00:42:22,000

The difference is essentially cost. You could think of it as revenues, but the other side of revenues is cost.

差異本質上是成本。你可以把它想成收益，但收益的另一面是成本。

322

00:42:22,000 --> 00:42:29,000

If the life of the asset is long, the TCO is low. This is the difference.

如果資產壽命長，總持有成本（TCO）就低。這就是差異所在。

323

00:42:29,000 --> 00:42:47,000

This is what it looks like when compute. The more you buy, the more you make.

這就是算力的樣子。買得越多，賺得越多。

324

00:42:47,000 --> 00:42:57,000

Now, all of you are experiencing this with me, isn't that right? All of your demand, your factories are working so hard.

現在，你們所有人都和我一起體驗著這一切，不是嗎？你們所有的需求，你們的工廠都在拼命工作。

325

00:42:57,000 --> 00:43:04,000

People are working so hard all across Taiwan because everybody wants to make money.

整個台灣的人都在拼命工作，因為每個人都想賺錢。

326

00:43:04,000 --> 00:43:12,000

They realize that AI, useful AI is here. Profitable AI is here.

他們意識到 AI，真正有用的 AI 已在眼前。有利可圖的 AI 已在眼前。

327

00:43:12,000 --> 00:43:18,000

Compute demand is incredibly high and compute demand is the constraint.

算力需求極高，而算力需求正是瓶頸所在。

328

00:43:18,000 --> 00:43:25,000

And so let's go work super, super hard and help the world stand up AI factories everywhere.

所以讓我們超級、超級努力工作，幫助全世界到處建立 AI 工廠。

329

00:43:25,000 --> 00:43:39,000

This is why it's so important. I'm so happy. Here I am standing in front of you. Vera Rubin is in full production.

這就是為什麼如此重要。我非常高興。我站在你們面前。Vera Rubin 已正式全面量產。

330

00:43:39,000 --> 00:43:43,000

Vera Rubin is in full production.

Vera Rubin 已正式全面量產。

331

00:43:43,000 --> 00:43:52,000

The supply chain we created for Vera Rubin is twice as large as Grace Blackwell.

我們為 Vera Rubin 創建的供應鏈規模是 Grace Blackwell 的兩倍。

332

00:43:52,000 --> 00:44:04,000

Not. Yeah, it's incredible. And and what used to take two hours to assemble one Grace Blackwell rack now only takes five minutes.

不，真的，難以置信。過去組裝一個 Grace Blackwell 機架需要兩個小時，現在只需要五分鐘。

333

00:44:04,000 --> 00:44:13,000

So not only is the capacity higher, the throughput is a lot faster and we need it all to support the demand.

所以不僅產能更高，生產速度也快得多，而我們需要所有這一切來支撐需求。

334

00:44:13,000 --> 00:44:25,000

This ecosystem is extraordinary. Millions of square feet has been put online to support Grace Blackwell and preparing now ramping up now Vera Rubin.

這個生態系統非常了不起。數百萬平方英尺的廠房已上線支援 Grace Blackwell，現在正在準備、加速推進 Vera Rubin。

335

00:44:25,000 --> 00:44:33,000

I want to thank all of you. Vera Rubin is now in full production. Thank you.

我要感謝你們所有人。Vera Rubin 現已全面量產。謝謝你們。

336

00:44:33,000 --> 00:44:37,000

Let's take a look.

讓我們來看看。

337

00:44:37,000 --> 00:44:41,000

Large language models generate answers.

大型語言模型生成答案。

338

00:44:41,000 --> 00:44:48,000

Now, AI agents can do work, but processing agentic AI is a whole different kind of problem.

現在，AI Agent 能夠完成工作，但處理 Agentic AI 是一種完全不同的問題。

339

00:44:48,000 --> 00:44:52,000

Agents observe reason plan use tools.

Agent 觀察、推理、規劃、使用工具。

340

00:44:52,000 --> 00:44:56,000

They manage massive context, juggling working memory and long term memory.

它們管理龐大的上下文，同時應對工作記憶和長期記憶。

341

00:44:56,000 --> 00:45:00,000

They spin up sub agents specialists on demand.

它們按需啟動子代理（sub-agent）專家。

342

00:45:00,000 --> 00:45:08,000

NVIDIA Vera Rubin is a multi rack pod scale system built to process agentic AI and is now in full production.

NVIDIA Vera Rubin 是一個多機架 Pod 規模系統，專為處理 Agentic AI 而建造，現已全面量產。

343

00:45:08,000 --> 00:45:12,000

The manufacturing automation and orchestration across the supply chain.

整個供應鏈的製造自動化和協調。

344

00:45:12,000 --> 00:45:20,000

A miracle to witness. Our journey started when we launched the first AI supercomputer NVIDIA DGX one.

見證奇蹟。我們的旅程始於推出第一台 AI 超級電腦 NVIDIA DGX One。

345

00:45:20,000 --> 00:45:31,000

Over the next decade, we pushed every chip and system to the limit from Pascal and the first and feeling to Grace Blackwell, the first rack scale AI supercomputer.

在接下來的十年裡，我們把每一顆晶片和系統推向極限，從 Pascal 世代到 Grace Blackwell，第一個機架規模的 AI 超級電腦。

346

00:45:31,000 --> 00:45:38,000

And now Vera Rubin, the first multi rack pod scale supercomputer built for the agentic age.

而現在是 Vera Rubin，第一台為 Agentic 時代建造的多機架 Pod 規模超級電腦。

347

00:45:38,000 --> 00:45:45,000

It starts at TSMC. The seven new chips that make up Vera Rubin take shape through hundreds of processing steps.

一切從台積電開始。構成 Vera Rubin 的七顆新晶片，經過數百個製程步驟成形。

348

00:45:45,000 --> 00:45:54,000

Three nanometer process, CoWAS R and CoWAS L packaging, HBM for memory from Micron, SK Hynix and Samsung.

3 奈米製程、CoWOS-R 和 CoWOS-L 封裝，來自美光、SK Hynix 和三星的 HBM 記憶體。

349

00:45:54,000 --> 00:46:02,000

The Vera Rubin Compute Board, six trillion transistors with over 18,000 components on one board.

Vera Rubin 運算板，一塊板子上有 6 兆個電晶體和超過 18,000 個元件。

350

00:46:02,000 --> 00:46:09,000

Vera Rubin MVL 72 does the thinking, prompt and context understanding, reasoning and planning.

Vera Rubin NVL 72 負責思考、提示和上下文理解、推理和規劃。

351

00:46:09,000 --> 00:46:15,000

Next, a new modular computer stream line with a new PCB mid plane design.

接著，一種採用全新 PCB 中板設計的新模組化電腦流線型設計。

352

00:46:15,000 --> 00:46:26,000

Superchips, Connect X9 SuperNix and Bluefield 4 DPUs all made in place with no cables for resiliency at AI factory scale.

Superchip、ConnectX-9、SuperNICS 和 Bluefield 4 DPU 全部就地安裝，無需電纜，在 AI 工廠規模下確保韌性。

353

00:46:26,000 --> 00:46:37,000

18 compute trays, nine hot swappable MVLink switch trays, new high efficiency manifolds, liquid cooled bus bars carrying over 5000 amps.

18 個運算托盤、9 個熱插拔 NVLink 交換機托盤、新型高效集管、液冷匯流排承載超過 5,000 安培。

354

00:46:37,000 --> 00:46:41,000

The equivalent of 20 electric cars at full acceleration.

相當於 20 輛電動車全力加速所需的電力。

355

00:46:41,000 --> 00:46:47,000

Together, 1.3 million components form this third generation MGX rack design.

合計，130 萬個元件構成了這個第三代 MGX 機架設計。

356

00:46:47,000 --> 00:46:53,000

Congratulations to Microsoft for their operational Vera Rubin MVL 72 engineering rack.

恭賀微軟擁有可運行的 Vera Rubin NVL 72 工程機架。

357

00:46:53,000 --> 00:47:00,000

Congratulations to Dell and CoreWeave as well for standing up their Vera Rubin MVL 72 engineering rack.

同時恭賀 Dell 和 CoreWeave 也建立了各自的 Vera Rubin NVL 72 工程機架。

358

00:47:00,000 --> 00:47:12,000

Then, the Vera CPU rack. 256 CPUs in a single liquid cooled rack, orchestrating the models, shuffling memory, launching tools.

接著是 Vera CPU 機架。一個液冷機架中有 256 顆 CPU，協調模型、調度記憶體、啟動工具。

359

00:47:12,000 --> 00:47:16,000

At Foxconn and Quanta, Grok 3 LPX takes shape.

在鴻海和廣達，Groq 3 LPX 成形。

360

00:47:16,000 --> 00:47:26,000

256 Grok 3 LPUs across 16 trays, 40 petabytes per second of SRAM bandwidth for ultra-low latency.

16 個托盤共 256 顆 Groq 3 LPU，40 PB/s 的 SRAM 頻寬，實現超低延遲。

361

00:47:26,000 --> 00:47:34,000

While MVL 72 generates tokens at the highest throughput, Grok LPX generates them at the lowest latency.

NVL 72 以最高吞吐量生成 token，而 Groq LPX 以最低延遲生成 token。

362

00:47:34,000 --> 00:47:39,000

Vera Bluefield 4 STX, where AI keeps its memory.

Vera Bluefield 4 STX，AI 儲存記憶體的地方。

363

00:47:39,000 --> 00:47:47,000

Storage processing accelerated by Bluefield 4, connecting memory, storage and in-silicon security.

由 Bluefield 4 加速的存儲處理，連接記憶體、存儲和晶片內安全功能。

364

00:47:47,000 --> 00:47:56,000

And NVIDIA Spectrum X Ethernet Photonics, the world's first Ethernet switch with 200 gigabit co-packaged optics.

還有 NVIDIA Spectrum X 乙太網光子技術，全球首款採用 200Gb 共封裝光學（CPO）的乙太網交換機。

365

00:47:56,000 --> 00:48:04,000

TSMC's Coupe Process, chip scale packaging and ultra-high powered laser dies on indium phosphide.

台積電的 CoUPE 製程、晶片級封裝和磷化銦上的超高功率雷射晶粒。

366

00:48:04,000 --> 00:48:11,000

Vera Rubin, five connected rack scale systems, a supercomputer for AI agents.

Vera Rubin，五個互連的機架規模系統，一台為 AI Agent 打造的超級電腦。

367

00:48:11,000 --> 00:48:17,000

150 supply chain partners across Taiwan. Millions of square feet of factory floor.

台灣各地 150 家供應鏈夥伴。數百萬平方英尺的工廠廠房。

368

00:48:17,000 --> 00:48:26,000

Hundreds of sites, chips, packages, systems and data centers pushed to the limits of size, power and scale.

數百個地點，晶片、封裝、系統和資料中心被推向尺寸、電力和規模的極限。

369

00:48:26,000 --> 00:48:30,000

This is what we call extreme co-design. We did this with Taiwan.

這就是我們所說的極致協同設計（extreme co-design）。我們與台灣一起做到了。

370

00:48:30,000 --> 00:48:34,000

Together, we reinvented computing for the age of AI.

我們攜手重新發明了 AI 時代的運算。

371

00:48:34,000 --> 00:48:40,000

Taiwan was with us at the beginning and here today as we bring Vera Rubin to the world.

台灣從一開始就與我們同在，今天也在這裡，我們將 Vera Rubin 帶給全世界。

372

00:48:40,000 --> 00:48:42,000

Thank you Taiwan.

謝謝台灣。

373

00:48:42,000 --> 00:48:52,000

Ladies and gentlemen, Vera Rubin.

女士們先生們，Vera Rubin。

374

00:48:52,000 --> 00:48:58,000

Vera Rubin was not just built for AI.

Vera Rubin 不只是為 AI 而建造的。

375

00:48:58,000 --> 00:49:01,000

Vera Rubin was not built just to run AI.

Vera Rubin 不只是為了運行 AI 而建造的。

376

00:49:01,000 --> 00:49:05,000

Vera Rubin was built to run agents.

Vera Rubin 是為了運行 Agent 而建造的。

377

00:49:05,000 --> 00:49:08,000

This is an agentic system.

這是一個 Agentic 系統。

378

00:49:08,000 --> 00:49:16,000

Imagine the complexity, which is the reason why agents is the last computer science breakthrough.

想像一下複雜度，這就是為什麼 Agent 是電腦科學最後一個重大突破的原因。

379

00:49:16,000 --> 00:49:21,000

It has taken this many years for agents to realize its potential and become useful.

Agent 花了這麼多年才實現其潛力並成為有用的工具。

380

00:49:21,000 --> 00:49:26,000

It stands to reason that the computer that runs it is the most advanced in the world.

能夠運行它的電腦自然是世界上最先進的。

381

00:49:26,000 --> 00:49:28,000

This is Vera Rubin.

這就是 Vera Rubin。

382

00:49:28,000 --> 00:49:31,000

Let's take a look. Can we bring out Vera Rubin, please?

讓我們來看看。能把 Vera Rubin 帶出來嗎？

383

00:49:50,000 --> 00:49:55,000

And Janine, do we have the racks, the systems?

還有 Janine，我們有機架和系統嗎？

384

00:49:59,000 --> 00:50:01,000

It looks heavy.

看起來很重。

385

00:50:01,000 --> 00:50:05,000

This is Vera Rubin.

這就是 Vera Rubin。

386

00:50:05,000 --> 00:50:08,000

Vera Rubin MVLink 72.

Vera Rubin NVLink 72。

387

00:50:08,000 --> 00:50:11,000

This is the Grok LPX.

這是 Groq LPX。

388

00:50:11,000 --> 00:50:14,000

At the next GTC, I'm going to talk to you about a lot more of this.

在下一次 GTC，我將向你們詳細介紹更多這方面的內容。

389

00:50:14,000 --> 00:50:17,000

Today, we have so much to talk to you about.

今天，我們有太多要和你們談的。

390

00:50:17,000 --> 00:50:20,000

This is Vera CPU rack.

這是 Vera CPU 機架。

391

00:50:20,000 --> 00:50:24,000

256 CPUs, all liquid cooled.

256 顆 CPU，全部液冷。

392

00:50:24,000 --> 00:50:26,000

Let me tell you about Vera in just a moment.

讓我稍後告訴你們關於 Vera 的事。

393

00:50:26,000 --> 00:50:34,000

This is the Vera Bluefield storage processing system and also security system.

這是 Vera Bluefield 存儲處理系統，也是安全系統。

394

00:50:34,000 --> 00:50:41,000

And of course, this is our Mellanox networking, the world's first CPO.

當然，這是我們的 Mellanox 網路，全球首款 CPO（共封裝光學）。

395

00:50:41,000 --> 00:50:42,000

This is Vera Rubin.

這就是 Vera Rubin。

396

00:50:42,000 --> 00:50:45,000

Incredible technology all coming together.

令人難以置信的技術匯聚在一起。

397

00:50:45,000 --> 00:50:52,000

Now, when we built Harper, we built Harper, as you know, for pre-training.

現在，當我們建造 Hopper 時，如你們所知，我們是為了預訓練（pre-training）而建造的。

398

00:50:52,000 --> 00:50:55,000

Pre-training was the most important application, the most important workload.

預訓練是最重要的應用、最重要的工作負載。

399

00:50:55,000 --> 00:50:57,000

We were working on at the time.

那是我們當時正在處理的。

400

00:50:57,000 --> 00:51:02,000

Then when we worked on Grace Blackwell, everybody said, Jensen, you know,

然後當我們在開發 Grace Blackwell 時，每個人都說：「Jensen，你知道，」

401

00:51:02,000 --> 00:51:05,000

Nvidia is really good at pre-training.

NVIDIA 在預訓練方面真的很厲害。

402

00:51:05,000 --> 00:51:07,000

Inference is so easy.

推論（inference）很簡單啊。

403

00:51:07,000 --> 00:51:09,000

Do you remember that?

你們還記得嗎？

404

00:51:09,000 --> 00:51:11,000

People used to say inference is so easy.

人們過去常說推論很簡單。

405

00:51:11,000 --> 00:51:13,000

We could do that too.

我們也能做到的。

406

00:51:13,000 --> 00:51:16,000

But as you know, inference equals money.

但如你們所知，推論等於錢。

407

00:51:16,000 --> 00:51:20,000

And the models, MOEs are so complicated.

而且這些模型，MoE（混合專家模型）是如此複雜。

408

00:51:20,000 --> 00:51:26,000

And to do it at incredibly high response time, fast interactivity,

要在極快的回應時間、快速的互動性，

409

00:51:26,000 --> 00:51:30,000

and high throughput at the same time is incredibly hard,

以及高吞吐量的情況下同時做到，是極其困難的，

410

00:51:30,000 --> 00:51:33,000

which is the reason why we created NVLink 72.

這就是我們創造 NVLink 72 的原因。

411

00:51:33,000 --> 00:51:40,000

Today, Nvidia's token cost is the lowest in the world, not by 10%,

今天，NVIDIA 的 token 成本是全球最低的，不是低 10%，

412

00:51:40,000 --> 00:51:46,000

by X factors, orders of magnitude, all because we did extreme co-design,

而是低幾倍、數個數量級，全因為我們進行了極致協同設計，

413

00:51:46,000 --> 00:51:52,000

all because we understood the computing model, the computing pattern of inference,

全因為我們理解了推論的運算模型和運算模式，

414

00:51:52,000 --> 00:51:55,000

and we were able to create NVLink 72.

並且能夠創造出 NVLink 72。

415

00:51:55,000 --> 00:52:00,000

Now, with Vera Rubin, it is beyond inference.

現在，有了 Vera Rubin，它超越了推論。

416

00:52:00,000 --> 00:52:04,000

It is now inference in an agentic system.

現在是在 Agentic 系統中的推論。

417

00:52:04,000 --> 00:52:07,000

This is Vera Rubin.

這就是 Vera Rubin。

418

00:52:07,000 --> 00:52:13,000

No cables, no hoses, no fans.

沒有電纜，沒有水管，沒有風扇。

419

00:52:13,000 --> 00:52:19,000

What used to take the last time when I showed this to you, we had cables everywhere.

上次我向你們展示時，到處都是電纜。

420

00:52:19,000 --> 00:52:22,000

The cables were amazing to look at.

那些電纜看起來令人嘆為觀止。

421

00:52:22,000 --> 00:52:27,000

But now, there's a PCB in the middle, which connects both sides.

但現在，中間有一塊 PCB，將兩側連接起來。

422

00:52:27,000 --> 00:52:31,000

What used to take two hours now takes five minutes.

過去需要兩個小時的工作，現在只需五分鐘。

423

00:52:31,000 --> 00:52:37,000

The reliability and the resilience of Vera Rubin is going to be off the charts.

Vera Rubin 的可靠性和韌性將超乎想像。

424

00:52:37,000 --> 00:52:40,000

This is our Vera CPU tray.

這是我們的 Vera CPU 托盤。

425

00:52:40,000 --> 00:52:44,000

The most advanced CPUs that has ever been built.

有史以來建造過的最先進 CPU。

426

00:52:44,000 --> 00:52:47,000

I'm going to show you that in just a second.

我馬上就要向你們展示。

427

00:52:47,000 --> 00:52:51,000

And this is our storage tray.

這是我們的存儲托盤。

428

00:52:51,000 --> 00:52:59,000

Two Vera CPUs, four CX-9, incredible amounts of software.

兩顆 Vera CPU、四個 CX-9，以及令人難以置信的大量軟體。

429

00:52:59,000 --> 00:53:08,000

This is our new LPX, LPU-30, the Grok system, designed for very low latency inference.

這是我們新的 LPX、LPU-30，Groq 系統，專為超低延遲推論設計。

430

00:53:08,000 --> 00:53:15,000

The throughput is delivered by Vera Rubin and extended with NVLink 72.

吞吐量由 Vera Rubin 提供，並通過 NVLink 72 擴展。

431

00:53:15,000 --> 00:53:21,000

If you want to extend that even further, you can have Grok LPUs.

如果你想進一步擴展，可以使用 Groq LPU。

432

00:53:21,000 --> 00:53:25,000

Here, we have the Vera Rubin NVLink, the switch tray.

這裡，我們有 Vera Rubin NVLink 交換機托盤。

433

00:53:25,000 --> 00:53:30,000

This is the switches in the middle, and this is revolutionary.

這是中間的交換機，這是革命性的。

434

00:53:30,000 --> 00:53:37,000

Because of Vera Rubin, because of NVLink 72 and the NVLink switches that we created and invented,

因為 Vera Rubin，因為我們創造和發明的 NVLink 72 和 NVLink 交換機，

435

00:53:37,000 --> 00:53:43,000

and this is our Ethernet switches for scale out.

這是我們用於橫向擴展的乙太網交換機。

436

00:53:43,000 --> 00:53:50,000

What's amazing is we introduced these two systems for Grace Blackwell.

令人驚嘆的是，我們為 Grace Blackwell 推出了這兩個系統。

437

00:53:50,000 --> 00:53:53,000

These two systems were created for Grace Blackwell,

這兩個系統是為 Grace Blackwell 創建的，

438

00:53:53,000 --> 00:53:58,000

and today, NVIDIA is the largest networking company in the world.

而今天，NVIDIA 是全球最大的網路公司。

439

00:53:58,000 --> 00:54:01,000

I'm so proud of the networking team.

我為網路團隊感到非常自豪。

440

00:54:01,000 --> 00:54:05,000

This is such an incredible enabler for everything that we do.

這對我們所做的一切都是如此重要的推動力。

441

00:54:05,000 --> 00:54:12,000

I'm going to now talk to you about the next major industry we're going to be part of.

我現在要和你們談談我們即將進入的下一個重要產業。

442

00:54:12,000 --> 00:54:13,000

Thank you.

謝謝。

443

00:54:13,000 --> 00:54:15,000

Janine?

Janine？

444

00:54:15,000 --> 00:54:17,000

Thank you.

謝謝。

445

00:54:22,000 --> 00:54:27,000

[speaking in Chinese]

【說中文】

446

00:54:27,000 --> 00:54:36,000

I think there are 2,000 people back there pulling that.

我想後面有 2,000 人在拉那個（機架）。

447

00:54:36,000 --> 00:54:41,000

Okay, let's talk about CPUs.

好的，讓我們談談 CPU。

448

00:54:41,000 --> 00:54:46,000

Vera CPUs, CPUs built for the age of AI.

Vera CPU，為 AI 時代而建造的 CPU。

449

00:54:46,000 --> 00:54:52,000

All of the CPUs until now were created for people.

到目前為止所有的 CPU 都是為人類創建的。

450

00:54:52,000 --> 00:54:54,000

We were the users.

我們是使用者。

451

00:54:54,000 --> 00:54:56,000

We were the users.

我們是使用者。

452

00:54:56,000 --> 00:54:58,000

We were the renters.

我們是租用者。

453

00:54:58,000 --> 00:55:04,000

The way we use CPUs, we live in a world counted by seconds.

我們使用 CPU 的方式，我們生活在一個以秒計時的世界裡。

454

00:55:04,000 --> 00:55:07,000

The way we rent CPUs in the cloud,

我們在雲端租用 CPU 的方式，

455

00:55:07,000 --> 00:55:13,000

each one of them, the more CPU cores you have, the more you can rent.

每一個核心，你有越多 CPU 核心，你能租出的就越多。

456

00:55:13,000 --> 00:55:19,000

The use case of the old CPU and the economics of the old CPU,

舊 CPU 的使用案例和舊 CPU 的經濟模型，

457

00:55:19,000 --> 00:55:22,000

fundamentally different than agents.

與 Agent 根本不同。

458

00:55:22,000 --> 00:55:25,000

Agents are impatient.

Agent 是沒有耐性的。

459

00:55:25,000 --> 00:55:27,000

They don't live in a world that is in seconds.

它們不生活在以秒計時的世界裡。

460

00:55:27,000 --> 00:55:30,000

They live in a world that's in nanoseconds.

它們生活在以奈秒計時的世界裡。

461

00:55:30,000 --> 00:55:36,000

When it uses a tool, it wants the response time to be as fast as possible.

當它使用工具時，它希望回應時間越快越好。

462

00:55:36,000 --> 00:55:41,000

When it accesses database, it has to come back as soon as possible.

當它存取資料庫時，必須盡快返回結果。

463

00:55:41,000 --> 00:55:49,000

Every moment that the agent is waiting keeps it from going to the next step, the next step, the next step.

Agent 每一刻的等待都阻礙它進入下一步、下一步、下一步。

464

00:55:49,000 --> 00:55:59,000

It is vital that we make the CPUs as low latency as possible, as interactive as possible.

使 CPU 的延遲盡可能低、互動性盡可能高，至關重要。

465

00:55:59,000 --> 00:56:03,000

So we created Vera CPU for the age of AI.

所以我們為 AI 時代創造了 Vera CPU。

466

00:56:03,000 --> 00:56:07,000

Now, inside our system, it's used for three different ways.

現在，在我們的系統中，它有三種不同的用途。

467

00:56:07,000 --> 00:56:14,000

The first way, of course, is Vera Rubin for thinking.

第一種用途當然是用於 Vera Rubin 的思考。

468

00:56:14,000 --> 00:56:19,000

And inside the Vera Rubin rack, there are already two CPUs.

而在 Vera Rubin 機架內，已經有兩顆 CPU。

469

00:56:19,000 --> 00:56:26,000

As you know, we are building and selling millions of Vera Rubins.

如你們所知，我們正在建造和銷售數百萬台 Vera Rubin。

470

00:56:26,000 --> 00:56:29,000

We have sold millions of Grace Blackwalls.

我們已經銷售了數百萬台 Grace Blackwell。

471

00:56:29,000 --> 00:56:33,000

NVIDIA already is one of the largest CPU makers in the world.

NVIDIA 已經是全球最大的 CPU 製造商之一。

472

00:56:33,000 --> 00:56:37,000

In the Vera Rubin rack are two CPUs.

Vera Rubin 機架中有兩顆 CPU。

473

00:56:37,000 --> 00:56:49,000

One for orchestrating and managing the GPUs, managing the KV cache, dealing with all of the software that runs in the rack.

一顆負責協調和管理 GPU、管理 KV 快取、處理機架中運行的所有軟體。

474

00:56:49,000 --> 00:56:56,000

We also have the Grace Bluefield that is used for security and isolation.

我們還有用於安全和隔離的 Grace Bluefield。

475

00:56:56,000 --> 00:57:06,000

The Vera compute is used for the harness, the orchestration of the AI models, tool use, accessing the database.

Vera 運算用於 harness、AI 模型的協調、工具使用、存取資料庫。

476

00:57:06,000 --> 00:57:18,000

And the data servers are right here, Vera Bluefield, the fastest storage servers, the fastest storage system the world has ever made.

而資料伺服器就在這裡，Vera Bluefield，有史以來最快的存儲伺服器、最快的存儲系統。

477

00:57:18,000 --> 00:57:26,000

And the reason why this is so vital is because agents are accessing memory so incredibly fast.

這如此至關重要的原因，是因為 Agent 以極其驚人的速度存取記憶體。

478

00:57:26,000 --> 00:57:38,000

These systems, the storage server and the CPUs, are now the critical path of the most expensive part of the data center.

這些系統——存儲伺服器和 CPU，現在是資料中心最昂貴部分的關鍵路徑。

479

00:57:38,000 --> 00:57:42,000

This is the most expensive for a good reason.

這是最昂貴的，有充分的理由。

480

00:57:42,000 --> 00:57:49,000

The economics of the AI factory is tokens.

AI 工廠的經濟模式就是 token。

481

00:57:49,000 --> 00:57:52,000

And the tokens are created here.

而 token 就在這裡被創造出來。

482

00:57:52,000 --> 00:57:57,000

And so of course you want to manufacture and generate as many tokens as possible.

所以當然你希望製造和生成盡可能多的 token。

483

00:57:57,000 --> 00:58:00,000

This is where you put all of your economics.

這就是你投入所有經濟資源的地方。

484

00:58:00,000 --> 00:58:03,000

And this has to not be in the way.

而這個（CPU）不能成為瓶頸。

485

00:58:03,000 --> 00:58:13,000

And so Vera's CPU has great pressure on the CPU architecture, which is the reason why we built a brand new architecture from the ground up.

所以 Vera 的 CPU 對 CPU 架構施加了巨大壓力，這就是我們從頭打造全新架構的原因。

486

00:58:13,000 --> 00:58:16,000

A CPU the world has never seen before.

一顆世界從未見過的 CPU。

487

00:58:16,000 --> 00:58:18,000

We call it Vera.

我們稱之為 Vera。

488

00:58:18,000 --> 00:58:22,000

This is CPU for agents.

這是為 Agent 而設計的 CPU。

489

00:58:22,000 --> 00:58:26,000

All the CPUs of the past we built for humans.

過去所有的 CPU 都是為人類建造的。

490

00:58:26,000 --> 00:58:29,000

This CPU is built for agents.

這顆 CPU 是為 Agent 建造的。

491

00:58:29,000 --> 00:58:31,000

Well, there are four things to keep in mind.

好，有四件事需要記住。

492

00:58:31,000 --> 00:58:33,000

The four takeaways.

四個重點。

493

00:58:33,000 --> 00:58:43,000

The first takeaway is that the instructions per clock of Vera has to be incredibly good because we need the latency to be short.

第一個重點是 Vera 的每時脈指令數（IPC）必須極其出色，因為我們需要延遲盡可能短。

494

00:58:43,000 --> 00:58:45,000

We need the processing time.

我們需要縮短處理時間。

495

00:58:45,000 --> 00:58:48,000

We need the single threaded performance.

我們需要單執行緒性能。

496

00:58:48,000 --> 00:58:49,000

Not throughput.

不是吞吐量。

497

00:58:49,000 --> 00:58:52,000

Single threaded performance has to be world class.

單執行緒性能必須是世界級的。

498

00:58:52,000 --> 00:58:54,000

Absolutely the best.

絕對是最好的。

499

00:58:54,000 --> 00:59:01,000

Single threaded performance, which is the reason why the IPC, the instructions per clock of Vera is so high.

單執行緒性能，這就是為什麼 Vera 的 IPC（每時脈指令數）如此之高。

500

00:59:01,000 --> 00:59:03,000

It's the highest in the world.

這是全球最高的。

501

00:59:03,000 --> 00:59:08,000

Ten instructions fetch, decoded and executed per clock.

每個時脈周期可擷取、解碼和執行十條指令。

502

00:59:08,000 --> 00:59:09,000

Number one.

第一點。

503

00:59:09,000 --> 00:59:11,000

Number two.

第二點。

504

00:59:11,000 --> 00:59:18,000

The bandwidth necessary to move data in and out for the CPU has to be utterly world class.

CPU 進出資料所需的頻寬必須絕對是世界級的。

505

00:59:18,000 --> 00:59:22,000

The second thing is bandwidth per core.

第二件事是每核心頻寬。

506

00:59:22,000 --> 00:59:25,000

The third is just bandwidth period.

第三是整體頻寬。

507

00:59:25,000 --> 00:59:26,000

We're moving.

我們在推進。

508

00:59:26,000 --> 00:59:37,000

Remember, I said earlier, agentic systems is fundamentally disaggregated and distributed, disaggregated and distributed.

記住，我之前說過，Agentic 系統從根本上是分解式和分散式的，分解式和分散式。

509

00:59:37,000 --> 00:59:43,000

When computing is disaggregated and distributed, networking becomes the problem.

當運算是分解式和分散式時，網路就成了關鍵問題。

510

00:59:43,000 --> 00:59:53,000

Therefore, we have to move the data around as fast as possible between the CPU cores and between the CPU and the storage, the CPU and the GPU.

因此，我們必須盡快在 CPU 核心之間、CPU 與存儲之間、CPU 與 GPU 之間移動資料。

511

00:59:53,000 --> 01:00:00,000

The bandwidth around the system and inside the CPU core has to be utterly world class.

系統周圍和 CPU 核心內部的頻寬必須絕對是世界級的。

512

01:00:00,000 --> 01:00:04,000

This is the first CPU that's been built a long time.

這是很長時間以來首顆以此標準建造的 CPU。

513

01:00:04,000 --> 01:00:10,000

That is literally at radical limits with a fabric that connects all of the CPU cores.

它真正處於極限邊界，採用連接所有 CPU 核心的互連結構（fabric）。

514

01:00:10,000 --> 01:00:12,000

That is speed of light.

這接近光速。

515

01:00:12,000 --> 01:00:16,000

Three point six terabytes per second.

每秒 3.6 TB。

516

01:00:16,000 --> 01:00:28,000

No triplet tax, no chip boundary crossings because we need to have everything because the CPU cores are talking to each other with extremely high bandwidth.

沒有三重稅（triplet tax），沒有晶片邊界交叉，因為我們需要讓一切順暢，CPU 核心之間以極高頻寬相互通信。

517

01:00:28,000 --> 01:00:31,000

They're not rented core per core per core.

它們不是按核心一個個出租的。

518

01:00:31,000 --> 01:00:33,000

They're all working together.

它們全部協同工作。

519

01:00:33,000 --> 01:00:36,000

The cross sectional bandwidth of Vera is off the charts.

Vera 的截面頻寬超乎想像。

520

01:00:36,000 --> 01:00:40,000

It's the first one to be PCI Express Gen six.

這是首顆採用 PCIe Gen 6 的 CPU。

521

01:00:40,000 --> 01:00:48,000

It is also the first one to have LPDDR5 with one point two terabytes per second.

這也是首顆採用 LPDDR5 的 CPU，達到每秒 1.2 TB 的頻寬。

522

01:00:48,000 --> 01:00:55,000

Three times two to three times the bandwidth of the highest performance CPUs on the outside.

外部頻寬是高性能 x86 CPU 的二到三倍。

523

01:00:55,000 --> 01:00:58,000

Three times the bandwidth on the inside.

內部頻寬也是三倍。

524

01:00:58,000 --> 01:01:03,000

The bandwidth per core and the bandwidth period is world class.

每核心頻寬和整體頻寬都是世界級的。

525

01:01:03,000 --> 01:01:06,000

Now, remember, I showed you earlier.

現在，記住，我之前展示過的。

526

01:01:06,000 --> 01:01:11,000

The number of CPU cores, the number of CPUs is going to be quite high.

CPU 核心的數量、CPU 的數量將會相當高。

527

01:01:11,000 --> 01:01:15,000

And the reason for that is very simple.

原因非常簡單。

528

01:01:15,000 --> 01:01:23,000

We created CPUs for humans in the past and humans, there are only one billion of us.

我們過去為人類建造 CPU，而人類只有 10 億。

529

01:01:23,000 --> 01:01:36,000

There will be billions of agents and these agents are going to be using the CPUs with very little patience because the cost of the GPU they sit next to is too high.

將會有數十億個 Agent，這些 Agent 將以極少的耐性使用 CPU，因為它們旁邊的 GPU 成本太高了。

530

01:01:36,000 --> 01:01:41,000

And therefore too valuable, too precious.

因此太有價值、太珍貴了。

531

01:01:41,000 --> 01:01:52,000

Therefore, these CPUs are going to be both performant, but they also have to be extremely energy efficient so that we can cram as much CPUs as possible.

因此，這些 CPU 既要性能強勁，也必須極其節能，這樣我們才能塞入盡可能多的 CPU。

532

01:01:52,000 --> 01:02:03,000

We can cram as much CPUs as we can into the factory without taking away power from the token generation, which we know is how we make money.

我們可以在工廠裡塞入盡可能多的 CPU，同時不搶走 token 生成所需的電力，那才是我們賺錢的方式。

533

01:02:03,000 --> 01:02:18,000

These four properties, instructions per clock or single threaded performance, bandwidth per core, the total bandwidth around the chip and inside the chip and energy efficiency defines Vera.

這四個特性——每時脈指令數（單執行緒性能）、每核心頻寬、晶片周圍和晶片內部的總頻寬，以及能源效率——定義了 Vera。

534

01:02:18,000 --> 01:02:20,000

It is absolutely world class.

它絕對是世界級的。

535

01:02:20,000 --> 01:02:25,000

When you compare it to the highest performance x86, it is just off the charts.

與最高性能的 x86 相比，它簡直超乎想像。

536

01:02:25,000 --> 01:02:34,000

When you compare it in real single threaded performance, real performance, it's off the charts.

在真實的單執行緒性能、真實性能比較中，它也是超乎想像的。

537

01:02:34,000 --> 01:02:39,000

It is incredible to be able to deliver 5% improvement on CPUs.

能夠在 CPU 上提供 5% 的改進已經令人難以置信。

538

01:02:39,000 --> 01:02:42,000

It is incredible to be able to deliver 10%.

能夠提供 10% 也令人難以置信。

539

01:02:42,000 --> 01:02:47,000

But this kind of performance speed up is just unheard of.

但這種程度的性能提升是前所未聞的。

540

01:02:47,000 --> 01:02:50,000

This is NVIDIA Vera.

這就是 NVIDIA Vera。

541

01:02:50,000 --> 01:02:57,000

What do you think?

你們覺得如何？

542

01:02:57,000 --> 01:03:00,000

Let's take a look.

讓我們來看看。

543

01:03:00,000 --> 01:03:03,000

Agentic AI changes the role of the CPU.

Agentic AI 改變了 CPU 的角色。

544

01:03:03,000 --> 01:03:07,000

The CPU is now the conductor and the GPU is the orchestra.

CPU 現在是指揮，GPU 是管弦樂團。

545

01:03:07,000 --> 01:03:13,000

Traditional CPUs were built for a different era, maximizing cores per socket.

傳統 CPU 是為另一個時代建造的，以每個插槽最大化核心數為目標。

546

01:03:13,000 --> 01:03:17,000

Slice them up, virtualize, rent by the hour.

切片、虛擬化、按小時出租。

547

01:03:17,000 --> 01:03:28,000

In the age of agents, the CPU is now a bottleneck to GPU utilization, directly affecting token throughput, latency and user experience.

在 Agent 時代，CPU 現在是 GPU 利用率的瓶頸，直接影響 token 吞吐量、延遲和使用者體驗。

548

01:03:28,000 --> 01:03:42,000

NVIDIA Vera is the CPU built for the agentic loop, combining NVIDIA's custom data center CPU core with a scalable coherency fabric for the right balance of performance cores and bandwidth to maximize AI factory output.

NVIDIA Vera 是為 Agentic 循環而建造的 CPU，結合 NVIDIA 定制的資料中心 CPU 核心和可擴展一致性結構，在性能核心和頻寬之間取得平衡，以最大化 AI 工廠輸出。

549

01:03:42,000 --> 01:03:54,000

At the heart of Vera is the NVIDIA Olympus Core, built for modern data center workloads, branch heavy Python runtimes, tool calls and sandbox code execution.

Vera 的核心是 NVIDIA Olympus Core，專為現代資料中心工作負載而建造，包括分支密集的 Python 執行環境、工具呼叫和沙箱程式碼執行。

550

01:03:54,000 --> 01:03:57,000

Each core is tuned for throughput.

每個核心都針對吞吐量進行了調整。

551

01:03:57,000 --> 01:04:01,000

A neural branch predictor evaluating two taken branches per cycle.

神經分支預測器每個週期評估兩個已採取的分支。

552

01:04:01,000 --> 01:04:05,000

A 10-wide decode engine brings in more work each cycle.

10 寬的解碼引擎每個週期帶入更多工作。

553

01:04:05,000 --> 01:04:08,000

A large out of order engine keeps instructions moving.

大型亂序引擎保持指令持續流動。

554

01:04:08,000 --> 01:04:13,000

Advanced prefetchers with a novel graph engine anticipating the next data path.

先進的預取器配備全新圖形引擎，預測下一條資料路徑。

555

01:04:13,000 --> 01:04:18,000

But fast cores only matter when data arrives correctly and on time.

但快速核心只有在資料準時正確到達時才有意義。

556

01:04:18,000 --> 01:04:28,000

Vera is the first CPU to use LPDDR5X memory, while correcting multiple errors simultaneously without compromising bandwidth.

Vera 是首顆使用 LPDDR5X 記憶體的 CPU，同時同步糾正多個錯誤而不影響頻寬。

557

01:04:28,000 --> 01:04:39,000

Vera achieves 40% lower peak memory latency versus x86, keeping cores fed on time through retrieval, analytics and sandbox execution.

Vera 實現了比 x86 低 40% 的峰值記憶體延遲，在檢索、分析和沙箱執行中保持核心及時供給。

558

01:04:39,000 --> 01:04:49,000

NVIDIA's second generation scalable coherency fabric unifies all 88 Olympus Cores on a monolithic mesh with separate dies for memory and I/O.

NVIDIA 第二代可擴展一致性結構將所有 88 個 Olympus Core 統一在單一網格上，記憶體和 I/O 使用獨立晶粒。

559

01:04:49,000 --> 01:04:58,000

Cores are not split across chiplets, enabling 50% faster core-to-core communication than traditional CPUs.

核心不跨小晶片（chiplet）分割，使核心間通信比傳統 CPU 快 50%。

560

01:04:58,000 --> 01:05:04,000

And memory coherent NVLink chip-to-chip connects GPUs directly to the fabric.

記憶體一致的 NVLink 晶片對晶片，將 GPU 直接連接到互連結構。

561

01:05:04,000 --> 01:05:13,000

Beyond GPUs, NVLink chip-to-chip can scale Vera up to multiple sockets, enabling massive bandwidth between CPUs.

超越 GPU，NVLink 晶片對晶片可以將 Vera 擴展到多個插槽，在 CPU 之間實現巨大頻寬。

562

01:05:13,000 --> 01:05:19,000

Vera delivers 1.8 times the agentic sandbox performance of x86 CPUs.

Vera 提供 x86 CPU 的 1.8 倍 Agentic 沙箱性能。

563

01:05:19,000 --> 01:05:25,000

Standalone Vera racks run agent sandboxes, tools, code and data pipelines.

獨立的 Vera 機架運行 Agent 沙箱、工具、程式碼和資料管道。

564

01:05:25,000 --> 01:05:31,000

Tightly coupled to Rubin GPUs, Vera keeps accelerated workflows moving.

與 Rubin GPU 緊密耦合，Vera 保持加速工作流程持續運轉。

565

01:05:31,000 --> 01:05:37,000

NVIDIA Vera BlueField 4 STX powers context memory and AI storage.

NVIDIA Vera BlueField 4 STX 為上下文記憶體和 AI 存儲提供動力。

566

01:05:37,000 --> 01:05:46,000

Compute Networking Storage Vera is the CPU for the age of agents.

運算、網路、存儲——Vera 是 Agent 時代的 CPU。

567

01:05:46,000 --> 01:05:55,000

This is going to be our new major growth driver.

這將是我們新的主要成長驅動力。

568

01:05:55,000 --> 01:06:01,000

The reviews are already coming out and it's pretty good.

評測已經開始出爐，結果相當不錯。

569

01:06:01,000 --> 01:06:03,000

That's pretty good stuff.

這真的很不錯。

570

01:06:10,000 --> 01:06:12,000

Now remember,

現在記住，

571

01:06:12,000 --> 01:06:21,000

Grace and Vera are also the most highly qualified CPUs in the world of AI

Grace 和 Vera 也是全球 AI 領域認證度最高的 CPU，

572

01:06:21,000 --> 01:06:26,000

because every single data center, every single cloud, every single enterprise,

因為每一個資料中心、每一個雲端、每一個企業，

573

01:06:26,000 --> 01:06:33,000

every company that works with NVIDIA on AI has already qualified.

每一家與 NVIDIA 合作 AI 的公司都已通過認證。

574

01:06:33,000 --> 01:06:37,000

Grace, the entire software stack has already been optimized for Grace.

Grace，整個軟體堆疊已針對 Grace 進行了優化。

575

01:06:37,000 --> 01:06:40,000

Every company will be qualifying Vera.

每家公司都將對 Vera 進行認證。

576

01:06:40,000 --> 01:06:45,000

Vera will be the most optimized agentic CPU in the world.

Vera 將是全球最優化的 Agentic CPU。

577

01:06:45,000 --> 01:06:48,000

Simply because it's going to go with Vera Rubin.

僅僅因為它將與 Vera Rubin 搭配。

578

01:06:48,000 --> 01:06:52,000

Simply because we made the big hard switch.

僅僅因為我們做出了那個艱難的重大切換。

579

01:06:52,000 --> 01:06:59,000

In fact, during Grace Blackwell's transition, the biggest risk was going from external CPU x86

事實上，在 Grace Blackwell 的轉型期間，最大的風險是從外部 x86 CPU

580

01:06:59,000 --> 01:07:01,000

into Grace Blackwell.

轉移到 Grace Blackwell。

581

01:07:01,000 --> 01:07:08,000

That transition was extremely dangerous, but we did it with incredible execution.

那個轉型極其危險，但我們以令人難以置信的執行力完成了它。

582

01:07:08,000 --> 01:07:13,000

Now Grace is literally synonymous with Grace Blackwell.

現在 Grace 字面上已與 Grace Blackwell 同義。

583

01:07:13,000 --> 01:07:15,000

When people say Blackwell, they say Grace Blackwell.

當人們說 Blackwell 時，他們說的是 Grace Blackwell。

584

01:07:15,000 --> 01:07:18,000

Because it is utterly now everywhere.

因為它現在無處不在。

585

01:07:18,000 --> 01:07:21,000

Every company's software stack has been optimized for it.

每家公司的軟體堆疊都已針對它進行了優化。

586

01:07:21,000 --> 01:07:24,000

Everybody's security stack has been optimized for it.

每個人的安全堆疊都已針對它進行了優化。

587

01:07:24,000 --> 01:07:26,000

And now here comes Vera.

現在 Vera 來了。

588

01:07:26,000 --> 01:07:27,000

I'm super excited about that.

我對此非常興奮。

589

01:07:27,000 --> 01:07:30,000

Now look at some of the performance numbers.

現在來看看一些性能數字。

590

01:07:30,000 --> 01:07:32,000

Speedup says one thing.

加速比說明了一件事。

591

01:07:32,000 --> 01:07:36,000

It is extremely hard to speed up SQL.

加速 SQL 是極其困難的。

592

01:07:36,000 --> 01:07:46,000

SQL, the most famous domain specific language, DSL, that has ever been created.

SQL，有史以來創造過最著名的領域特定語言（DSL）。

593

01:07:46,000 --> 01:07:51,000

Before SQL, before CUDA, there was SQL.

在 CUDA 之前，在 OpenGL 之前，有 SQL。

594

01:07:51,000 --> 01:07:54,000

Before OpenGL, there was SQL.

在 OpenGL 之前，有 SQL。

595

01:07:54,000 --> 01:07:56,000

Invented by IBM.

由 IBM 發明。

596

01:07:56,000 --> 01:08:00,000

Today it is the structured database engine of the planet.

今天它是全球的結構化資料庫引擎。

597

01:08:00,000 --> 01:08:02,000

Everybody uses SQL.

每個人都在使用 SQL。

598

01:08:02,000 --> 01:08:06,000

This is SQL running three times faster.

這是快了三倍的 SQL。

599

01:08:06,000 --> 01:08:09,000

Not 10% faster, not 25% faster.

不是快 10%，不是快 25%。

600

01:08:09,000 --> 01:08:12,000

Three times faster.

快了三倍。

601

01:08:12,000 --> 01:08:13,000

Incredible.

令人難以置信。

602

01:08:13,000 --> 01:08:19,000

This is real time, the next one is real time stream processing.

這是即時的，下一個是即時串流處理。

603

01:08:19,000 --> 01:08:24,000

Remember, your AI is going to be not just reading documents.

記住，你的 AI 不只是要閱讀文件。

604

01:08:24,000 --> 01:08:31,000

Your AI is going to be watching for telemetry, especially inside a factory, inside a stock exchange.

你的 AI 將監控遙測數據，特別是在工廠內、在股票交易所內。

605

01:08:31,000 --> 01:08:35,000

You're going to be looking for telemetry continuously.

你將持續監測遙測數據。

606

01:08:35,000 --> 01:08:39,000

The burst of data that's coming in goes into a CPU.

傳入的突發資料進入 CPU。

607

01:08:39,000 --> 01:08:46,000

This is Vera CPU running real time stream processing for New York Stock Exchange.

這是 Vera CPU 為紐約證券交易所運行即時串流處理。

608

01:08:46,000 --> 01:08:51,000

Lynn Martin, the president of New York Stock Exchange, has been so gracious to partner with us.

紐約證券交易所總裁 Lynn Martin 非常慷慨地與我們合作。

609

01:08:51,000 --> 01:08:57,000

This system is run all over the world in real time stream processing.

這個系統在全球各地進行即時串流處理。

610

01:08:57,000 --> 01:08:59,000

Vera CPU six times.

Vera CPU 快了六倍。

611

01:08:59,000 --> 01:09:01,000

All because of the bandwidth.

全都是因為頻寬。

612

01:09:01,000 --> 01:09:05,000

The single threaded instruction execution.

單執行緒指令執行。

613

01:09:05,000 --> 01:09:07,000

The bandwidth inside between the cores.

核心之間的內部頻寬。

614

01:09:07,000 --> 01:09:09,000

The bandwidth outside.

外部頻寬。

615

01:09:09,000 --> 01:09:12,000

Vera is completely revolutionary.

Vera 是完全革命性的。

616

01:09:12,000 --> 01:09:15,000

That's Vera.

這就是 Vera。

617

01:09:20,000 --> 01:09:24,000

You know, X-Factors is something you talk about when you're talking about GPUs.

你知道，X 倍速是談論 GPU 時才會提到的事。

618

01:09:24,000 --> 01:09:30,000

It is quite rare that somebody talks about X-Factors on real workload.

在真實工作負載上談論 X 倍速是相當罕見的。

619

01:09:30,000 --> 01:09:33,000

Real workload that is associated with CPUs.

與 CPU 相關的真實工作負載。

620

01:09:33,000 --> 01:09:35,000

I'm so proud of the team.

我為團隊感到非常自豪。

621

01:09:35,000 --> 01:09:36,000

You guys did such a great job.

你們做得非常出色。

622

01:09:36,000 --> 01:09:39,000

We have an extraordinary roadmap coming.

我們有一個非凡的路線圖即將到來。

623

01:09:39,000 --> 01:09:46,000

But what's really exciting is almost everybody is supporting Vera.

但真正令人興奮的是幾乎每個人都在支持 Vera。

624

01:09:46,000 --> 01:09:48,000

They're as excited as we are.

他們和我們一樣興奮。

625

01:09:48,000 --> 01:09:50,000

This is Vera opening up.

這就是 Vera 開啟的新局面。

626

01:09:50,000 --> 01:09:54,000

It's opened up a brand new market.

它開啟了一個全新的市場。

627

01:09:54,000 --> 01:09:55,000

Agents.

Agent。

628

01:09:55,000 --> 01:09:58,000

Agents is a new workload.

Agent 是一種新的工作負載。

629

01:09:58,000 --> 01:10:01,000

We built CPUs for humans in the past.

我們過去為人類建造 CPU。

630

01:10:01,000 --> 01:10:04,000

We need CPUs for agents.

我們需要為 Agent 而設計的 CPU。

631

01:10:04,000 --> 01:10:06,000

We need Agents for Agents.

我們需要為 Agent 服務的 CPU（Agent）。

632

01:10:06,000 --> 01:10:08,000

Agents are the new Agents.

Agent 是全新的使用者。

633

01:10:08,000 --> 01:10:10,000

Agents are the new Agents.

Agent 是全新的使用者。

634

01:10:10,000 --> 01:10:12,000

Agents are the new Agents.

Agent 是全新的使用者。

635

01:10:12,000 --> 01:10:14,000

Agents are the new Agents.

Agent 是全新的使用者。

636

01:10:14,000 --> 01:10:16,000

Agents are the new Agents.

Agent 是全新的使用者。

637

01:10:16,000 --> 01:10:18,000

Agents are the new Agents.

Agent 是全新的使用者。

638

01:10:18,000 --> 01:10:20,000

Agents are the new Agents.

Agent 是全新的使用者。

639

01:10:20,000 --> 01:10:22,000

Agents are the new Agents.

Agent 是全新的使用者。

640

01:10:22,000 --> 01:10:24,000

Agents are the new Agents.

Agent 是全新的使用者。

641

01:10:24,000 --> 01:10:26,000

Agents are the new Agents.

Agent 是全新的使用者。

642

01:10:26,000 --> 01:10:28,000

Agents are the new Agents.

Agent 是全新的使用者。

643

01:10:28,000 --> 01:10:30,000

Agents are the new Agents.

Agent 是全新的使用者。

644

01:10:30,000 --> 01:10:32,000

Agents are the new Agents.

Agent 是全新的使用者。

645

01:10:32,000 --> 01:10:34,000

Agents are the new Agents.

Agent 是全新的使用者。

646

01:10:34,000 --> 01:10:36,000

Agents are the new Agents.

Agent 是全新的使用者。

647

01:10:36,000 --> 01:10:38,000

Agents are the new Agents.

Agent 是全新的使用者。

648

01:10:38,000 --> 01:10:40,000

Agents are the new Agents.

Agent 是全新的使用者。

649

01:10:40,000 --> 01:10:42,000

Agents are the new Agents.

Agent 是全新的使用者。

650

01:10:42,000 --> 01:10:44,000

Agents are the new Agents.

Agent 是全新的使用者。

651

01:10:44,000 --> 01:10:46,000

Agents are the new Agents.

Agent 是全新的使用者。

652

01:10:46,000 --> 01:10:48,000

Agents are the new Agents.

Agent 是全新的使用者。

653

01:10:48,000 --> 01:10:50,000

Agents are the new Agents.

Agent 是全新的使用者。

654

01:10:50,000 --> 01:10:52,000

Agents are the new Agents.

Agent 是全新的使用者。

655

01:10:52,000 --> 01:10:54,000

Agents are the new Agents.

Agent 是全新的使用者。

656

01:10:54,000 --> 01:10:56,000

Agents are the new Agents.

Agent 是全新的使用者。

657

01:10:56,000 --> 01:10:58,000

Agents are the new Agents.

Agent 是全新的使用者。

658

01:10:58,000 --> 01:11:00,000

Agents are the new Agents.

Agent 是全新的使用者。

659

01:11:05,000 --> 01:11:07,000

This is the most important slide, really.

這真的是最重要的一張投影片。

660

01:11:07,000 --> 01:11:09,000

This is the takeaway.

這是重點所在。

661

01:11:09,000 --> 01:11:13,000

The takeaway here is that this is the application pattern.

這裡的重點是，這就是應用模式。

662

01:11:13,000 --> 01:11:17,000

This is the computing pattern of the next decade.

這是未來十年的運算模式。

663

01:11:17,000 --> 01:11:24,000

Agents, harnesses, orchestrating large language models.

Agent、harness，協調大型語言模型。

664

01:11:24,000 --> 01:11:27,000

Every company will run it.

每家公司都將運行它。

665

01:11:27,000 --> 01:11:30,000

Every company will be an Agent company.

每家公司都將成為 Agent 公司。

666

01:11:30,000 --> 01:11:34,000

Every company will have Agents running inside.

每家公司內部都將有 Agent 在運行。

667

01:11:34,000 --> 01:11:41,000

Every company will see that Agents will need its own operating system.

每家公司都將意識到，Agent 需要自己的作業系統。

668

01:11:41,000 --> 01:11:46,000

Every company is asking us, how do we run Agents safely?

每家公司都在問我們：我們如何安全地運行 Agent？

669

01:11:46,000 --> 01:11:50,000

How do we build Agents for our own workloads?

如何為我們自己的工作負載建構 Agent？

670

01:11:50,000 --> 01:11:56,000

And so we have the NVIDIA Agent Toolkit for Enterprise AI.

所以我們有了面向企業 AI 的 NVIDIA Agent Toolkit。

671

01:11:56,000 --> 01:11:59,000

You've seen me build this in plain sight.

你們已經見證了我公開地建構這個東西。

672

01:11:59,000 --> 01:12:02,000

Almost everything that NVIDIA does, as you know, at every GTC,

幾乎 NVIDIA 所做的一切，如你所知，在每一次 GTC，

673

01:12:02,000 --> 01:12:06,000

if you go back and look at my GTC five years ago or ten years ago,

如果你回顧我五年前或十年前的 GTC，

674

01:12:06,000 --> 01:12:08,000

you will see today.

你將看到今天的樣子。

675

01:12:08,000 --> 01:12:12,000

This you've seen me talking about for several years now

這個你已經看到我談論了好幾年了，

676

01:12:12,000 --> 01:12:15,000

because we've been building for this moment.

因為我們一直在為這一刻而建造。

677

01:12:15,000 --> 01:12:21,000

There are four things that companies need in order to build Agents as a service

公司需要四樣東西才能建構 Agent 即服務，

678

01:12:21,000 --> 01:12:24,000

or build Agents to operate.

或者建構 Agent 來操作。

679

01:12:24,000 --> 01:12:27,000

The first thing you need is you need models.

你需要的第一樣東西是模型。

680

01:12:27,000 --> 01:12:30,000

Of course, large language models.

當然，大型語言模型。

681

01:12:30,000 --> 01:12:34,000

The smarter the better, the cheaper the better, the faster the better.

越聰明越好，越便宜越好，越快越好。

682

01:12:34,000 --> 01:12:39,000

The second is you need a harness to orchestrate the whole thing.

第二是你需要一個 harness 來協調整個系統。

683

01:12:39,000 --> 01:12:46,000

The third, these models want to use tools, and these tools come with its skills.

第三，這些模型要使用工具，而這些工具附帶了技能（skills）。

684

01:12:46,000 --> 01:12:49,000

And I showed you CUDA X libraries.

我展示了 CUDA-X 函式庫。

685

01:12:49,000 --> 01:12:53,000

Those are going to be amazing tools for the Agents in the future.

那些將是未來 Agent 的絕佳工具。

686

01:12:53,000 --> 01:12:56,000

And then lastly, you need a runtime.

最後，你需要一個 runtime。

687

01:12:56,000 --> 01:12:59,000

You need the operating system that holds it all together.

你需要一個將所有東西整合在一起的作業系統。

688

01:12:59,000 --> 01:13:04,000

This is the NVIDIA Toolkit for Agents.

這就是 NVIDIA 的 Agent 工具套件。

689

01:13:04,000 --> 01:13:10,000

It includes models that you can modify.

它包含你可以修改的模型。

690

01:13:10,000 --> 01:13:14,000

NVIDIA's world-class open models, and I want to show you more.

NVIDIA 世界級的開放模型，我想向你們展示更多。

691

01:13:14,000 --> 01:13:18,000

You can run Agents from anybody.

你可以運行任何人的 Agent。

692

01:13:18,000 --> 01:13:23,000

You can run Cloud Code, Incredible Agent, Codex, Incredible Agent.

你可以運行 Claude Code（令人難以置信的 Agent）、Codex（令人難以置信的 Agent）。

693

01:13:23,000 --> 01:13:26,000

You can run it inside this harness called Open Shell,

你可以在這個稱為 Open Shell 的 harness 中運行它，

694

01:13:26,000 --> 01:13:30,000

which will be highly secure for you inside the enterprise.

它在企業內部對你而言將是高度安全的。

695

01:13:30,000 --> 01:13:38,000

The Shell protects the Agent, keeps it grounded in security policies.

Shell 保護 Agent，使其遵守安全政策。

696

01:13:38,000 --> 01:13:40,000

Privacy is protected.

隱私受到保護。

697

01:13:40,000 --> 01:13:43,000

Its rights and privileges are given.

其權利和特權得到授予。

698

01:13:43,000 --> 01:13:45,000

Its identity is protected.

其身份受到保護。

699

01:13:45,000 --> 01:13:49,000

And so this Open Shell is being adopted all over the world.

所以這個 Open Shell 正在全球各地被採用。

700

01:13:49,000 --> 01:13:51,000

NVIDIA Open Shell is open source.

NVIDIA Open Shell 是開源的。

701

01:13:51,000 --> 01:13:53,000

You're going to see so many companies adopt it.

你將看到許多公司採用它。

702

01:13:53,000 --> 01:13:59,000

Red Hat, Canonical, Microsoft, it's going to be adopted everywhere.

Red Hat、Canonical、Microsoft，它將無處不被採用。

703

01:13:59,000 --> 01:14:03,000

This is an important, this is the runtime.

這是重要的，這就是 runtime。

704

01:14:03,000 --> 01:14:08,000

And this runtime is fully optimized for the NVIDIA AI Platform,

而這個 runtime 完全針對 NVIDIA AI 平台進行了優化，

705

01:14:08,000 --> 01:14:09,000

which is everywhere.

而 NVIDIA AI 平台無處不在。

706

01:14:09,000 --> 01:14:16,000

So you can run Open Shell in any cloud, on-prem, and even on-device.

所以你可以在任何雲端、on-premise（本地端）甚至 on-device（裝置端）運行 Open Shell。

707

01:14:16,000 --> 01:14:21,000

So you have now tools and libraries that they can use.

所以你現在擁有工具和函式庫供 Agent 使用。

708

01:14:21,000 --> 01:14:25,000

You have models that you can modify or use as is.

你有可以修改或直接使用的模型。

709

01:14:25,000 --> 01:14:27,000

Or you have agents.

或者你有 Agent。

710

01:14:27,000 --> 01:14:34,000

This would be OpenClaw, Hermes, another incredible harness.

這包括 OpenClaw、Hermes，另一個令人難以置信的 harness。

711

01:14:34,000 --> 01:14:40,000

These agentic harnesses can now run on-prem or for you anywhere.

這些 Agentic harness 現在可以在本地端或任何地方為你運行。

712

01:14:40,000 --> 01:14:47,000

So four things, and this represents the operating system of the modern enterprise.

所以四樣東西，這代表了現代企業的作業系統。

713

01:14:47,000 --> 01:14:49,000

Now, how do we use this?

現在，我們如何使用它？

714

01:14:49,000 --> 01:14:55,000

One of my favorite use cases of agents is chip designers.

我最喜歡的 Agent 使用案例之一是晶片設計師。

715

01:14:55,000 --> 01:14:59,000

It is the single most important thing that NVIDIA does.

這是 NVIDIA 所做的最重要的事情。

716

01:14:59,000 --> 01:15:05,000

And so, of course, we have to partner with Cadence to build SuperAgent,

所以，當然，我們必須與 Cadence 合作建造 SuperAgent，

717

01:15:05,000 --> 01:15:08,000

a chip design SuperAgent.

一個晶片設計 SuperAgent。

718

01:15:08,000 --> 01:15:13,000

It is orchestrated by codex or cloud code.

它由 Codex 或 Claude Code 協調。

719

01:15:13,000 --> 01:15:20,000

It has RTL and architecture diagrams or schematics or specifications as input

它以 RTL 和架構圖、電路圖或規格說明作為輸入，

720

01:15:20,000 --> 01:15:22,000

and whatever you need to fix.

以及任何需要修正的東西。

721

01:15:22,000 --> 01:15:30,000

And together we created some SuperAgents that are optimized for the NVIDIA runtime

我們共同創建了針對 NVIDIA runtime 優化的 SuperAgent，

722

01:15:30,000 --> 01:15:34,000

with Nemotron, and let's take a look.

使用 Nemotron，讓我們來看看。

723

01:15:34,000 --> 01:15:36,000

It's really incredible.

真的令人難以置信。

724

01:15:36,000 --> 01:15:43,000

Cadence and NVIDIA are partnering to build chip design agents.

Cadence 和 NVIDIA 正在合作建造晶片設計 Agent。

725

01:15:43,000 --> 01:15:47,000

Hundreds of thousands of NVIDIA chips come together to make the AI factories

數十萬顆 NVIDIA 晶片共同構成 AI 工廠，

726

01:15:47,000 --> 01:15:51,000

that power the world's frontier AI models.

為世界最先進的 AI 模型提供動力。

727

01:15:51,000 --> 01:15:57,000

Finding these chips and the systems they run in is one of the hardest engineering challenges.

設計這些晶片和它們所在的系統是最困難的工程挑戰之一。

728

01:15:57,000 --> 01:16:03,000

Trillions of transistors, three-dimensional circuits, microscopic scale, every gate,

兆個電晶體、三維電路、微觀尺度，每一個邏輯閘，

729

01:16:03,000 --> 01:16:10,000

every wire, synchronized to picoseconds, must work in perfect harmony with no margin for error.

每一根連線，精確到皮秒，必須完美協調，沒有任何出錯的餘地。

730

01:16:10,000 --> 01:16:13,000

Physical prototypes are too slow and too costly.

實體原型太慢且成本太高。

731

01:16:13,000 --> 01:16:16,000

So engineers work in the digital realm.

所以工程師在數位領域工作。

732

01:16:16,000 --> 01:16:21,000

Each chip begins as a set of architectural specifications, then translated into RTL,

每顆晶片從一套架構規格開始，然後轉化為 RTL（暫存器傳輸層），

733

01:16:21,000 --> 01:16:24,000

the language of chip design.

晶片設計的語言。

734

01:16:24,000 --> 01:16:26,000

RTL must be verified in simulation.

RTL 必須通過模擬驗證。

735

01:16:26,000 --> 01:16:29,000

A single bug can delay a chip by months.

一個 bug 就可能讓晶片延遲數月。

736

01:16:29,000 --> 01:16:34,000

At NVIDIA, thousands of engineers, billions of compute hours per year,

在 NVIDIA，數千名工程師，每年數十億算力小時，

737

01:16:34,000 --> 01:16:39,000

millions of tests written, run, and debugged, a cycle that takes teams weeks.

數百萬個測試被編寫、執行和除錯，一個周期要花團隊數週時間。

738

01:16:39,000 --> 01:16:45,000

To compress this cycle, Cadence and NVIDIA built a design verification agent.

為了壓縮這個周期，Cadence 和 NVIDIA 建造了一個設計驗證 Agent。

739

01:16:45,000 --> 01:16:47,000

The text orchestrates the process.

文字（prompt）協調整個流程。

740

01:16:47,000 --> 01:16:51,000

Cadence chip stack launches the RTL verification loop,

Cadence 晶片堆疊啟動 RTL 驗證循環，

741

01:16:51,000 --> 01:16:55,000

由 Nemotron 驅動，由 NVIDIA Open Shell 保護，

742

01:16:55,000 --> 01:17:03,000

calling on expert sub-agents in RTL generation, test bench creation, regression testing, and debug.

呼叫 RTL 生成、測試台創建、回歸測試和除錯的專家子代理（sub-agent）。

743

01:17:03,000 --> 01:17:05,000

The system drives itself.

系統自我驅動。

744

01:17:05,000 --> 01:17:09,000

The chip stack agents run hundreds of simulations with Cadence Exelium,

晶片堆疊 Agent 使用 Cadence Exelium 運行數百次模擬，

745

01:17:09,000 --> 01:17:11,000

formal verification with Jasper.

使用 Jasper 進行形式驗證。

746

01:17:11,000 --> 01:17:15,000

Design flaws reveal, bugs in the code fixed.

設計缺陷被揭露，程式碼中的 bug 被修復。

747

01:17:15,000 --> 01:17:18,000

What once took weeks now takes hours.

過去需要數週的工作現在只需數小時。

748

01:17:18,000 --> 01:17:22,000

Verification cycles over 40 times faster.

驗證周期快了 40 倍以上。

749

01:17:22,000 --> 01:17:29,000

Together, NVIDIA and Cadence are reinventing chip design with AI agents.

共同地，NVIDIA 和 Cadence 正在用 AI Agent 重新發明晶片設計。

750

01:17:29,000 --> 01:17:36,000

From weeks to hours, from weeks to hours, from weeks to hours,

從數週到數小時，從數週到數小時，從數週到數小時，

751

01:17:36,000 --> 01:17:39,000

NVIDIA has thousands of chip designers.

NVIDIA 有數千名晶片設計師。

752

01:17:39,000 --> 01:17:45,000

We are going to hire hundreds of thousands of Cadence super agents that work with us,

我們將雇用數十萬個與我們合作的 Cadence SuperAgent，

753

01:17:45,000 --> 01:17:50,000

so that we can accelerate our company, so that we can be even more ambitious,

這樣我們就能加速公司發展，這樣我們就能更加雄心勃勃，

754

01:17:50,000 --> 01:17:54,000

create even more amazing things, run even faster.

創造出更驚人的事物，跑得更快。

755

01:17:54,000 --> 01:18:02,000

You saw earlier that the toolkit with models, harness, tools,

你之前看到工具套件有模型、harness、工具，

756

01:18:02,000 --> 01:18:07,000

the tools in this case are Cadence simulators and verifiers, formal verification systems.

在這個案例中，工具是 Cadence 的模擬器和驗證器、形式驗證系統。

757

01:18:07,000 --> 01:18:14,000

It is the reason why we are working with Cadence so hard to accelerate all of their tools on CUDA,

這就是我們為什麼努力與 Cadence 合作，在 CUDA 上加速他們所有工具的原因，

758

01:18:14,000 --> 01:18:16,000

because the agents are impatient.

因為 Agent 是沒有耐性的。

759

01:18:16,000 --> 01:18:19,000

The agents want the answer immediately.

Agent 希望立即得到答案。

760

01:18:19,000 --> 01:18:29,000

And so, models, harnesses, accelerated CUDA, accelerated libraries and tools, and then the runtime.

所以，模型、harness、加速的 CUDA、加速的函式庫和工具，然後是 runtime。

761

01:18:29,000 --> 01:18:33,000

What you saw just now is all of that coming together.

你剛才看到的就是所有這些匯聚在一起。

762

01:18:33,000 --> 01:18:41,000

Now, one of the things that it starts with is a great model that Cadence could modify and tune

現在，其中一件事是從一個 Cadence 可以修改和調整的優秀模型開始，

763

01:18:41,000 --> 01:18:45,000

to be expert at the Cadence workflow, at the Cadence expertise,

使其成為 Cadence 工作流程、Cadence 專業知識的專家，

764

01:18:45,000 --> 01:18:53,000

so that they could create super agents that are proprietary to Cadence with their proprietary knowledge.

這樣他們就能以自己的專有知識，創建 Cadence 專屬的 SuperAgent。

765

01:18:53,000 --> 01:18:57,000

They have to start with an excellent model. We call it Nemotron.

他們必須從一個出色的模型開始。我們稱之為 Nemotron。

766

01:18:57,000 --> 01:19:01,000

NVIDIA is dedicated to build open models for the world,

NVIDIA 致力於為全世界建造開放模型，

767

01:19:01,000 --> 01:19:06,000

so that all of you, all of us, could create our own agents.

這樣你們所有人、我們所有人，都能創建自己的 Agent。

768

01:19:06,000 --> 01:19:13,000

Today we're announcing the Nemotron 3 Ultra.

今天我們宣布 Nemotron 3 Ultra。

769

01:19:13,000 --> 01:19:18,000

Yep, our next open model. And it is smart.

對，我們的下一個開放模型。而且它很聰明。

770

01:19:18,000 --> 01:19:27,000

The Nemotron models not only give you the model,

Nemotron 模型不只給你模型，

771

01:19:27,000 --> 01:19:31,000

we give you all the data that we use to train the model,

我們給你我們用來訓練模型的所有資料，

772

01:19:31,000 --> 01:19:34,000

and because we have a coalition of incredible partners,

而且因為我們有一個了不起的合作夥伴聯盟，

773

01:19:34,000 --> 01:19:37,000

you can see all of our partners down here,

你可以在這裡看到我們所有的合作夥伴，

774

01:19:37,000 --> 01:19:41,000

we work together, contribute data to each other.

我們共同合作，互相貢獻資料。

775

01:19:41,000 --> 01:19:47,000

Nemotron is trained on one of the largest suites of long-running reasoning models,

Nemotron 在全球最大的長程推理模型訓練集之一上進行訓練，

776

01:19:47,000 --> 01:19:54,000

long-running tool task solving tool using data sets in the world,

全球最大的長程工具任務解決和工具使用資料集，

777

01:19:54,000 --> 01:19:57,000

because of all of our great partnerships.

這一切都因為我們出色的合作夥伴關係。

778

01:19:57,000 --> 01:20:05,000

All of this from the model, the training script, and the data made completely available to you.

從模型、訓練腳本到資料，這一切都完全向你開放。

779

01:20:05,000 --> 01:20:08,000

This is open models at its best.

這是最優質的開放模型。

780

01:20:08,000 --> 01:20:12,000

The best open model system policies in the world.

全球最佳的開放模型系統政策。

781

01:20:12,000 --> 01:20:19,000

Simple goal is so that you can take all of it, add to it, make it even better, make it yours.

簡單的目標：讓你能夠取用所有這些，加以增強，使其更好，使其成為你的。

782

01:20:19,000 --> 01:20:25,000

Nemotron 3 Ultra is five times faster.

Nemotron 3 Ultra 快了五倍。

783

01:20:25,000 --> 01:20:35,000

This is the world's first model based on a hybrid architecture of SSM state-space models with mixture of experts.

這是世界上第一個基於 SSM（狀態空間模型）與混合專家模型（MoE）混合架構的模型。

784

01:20:35,000 --> 01:20:38,000

The architecture is incredibly fast.

這個架構速度極快。

785

01:20:38,000 --> 01:20:40,000

We made it fast so that you could think fast.

我們讓它快，這樣你就能快速思考。

786

01:20:40,000 --> 01:20:44,000

When you think fast, you could think longer at the same cost.

當你思考得快，你就能以相同的成本思考更長時間。

787

01:20:44,000 --> 01:20:46,000

So five times faster.

所以快了五倍。

788

01:20:46,000 --> 01:20:50,000

It is also 30% cheaper.

它也便宜 30%。

789

01:20:50,000 --> 01:20:58,000

30% lower cost to run in total flops and total inference time than even the most cost effective in the world.

在總浮點運算量和總推論時間上，比全球最具成本效益的模型還要低 30% 的成本。

790

01:20:58,000 --> 01:21:01,000

We're comparing against the world's best open models.

我們是在與全球最佳開放模型進行比較。

791

01:21:01,000 --> 01:21:04,000

Frontier smart.

前沿智能。

792

01:21:04,000 --> 01:21:06,000

Five times faster.

五倍更快。

793

01:21:06,000 --> 01:21:08,000

30% cheaper.

便宜 30%。

794

01:21:08,000 --> 01:21:10,000

Completely open.

完全開放。

795

01:21:10,000 --> 01:21:12,000

We're completely dedicated to this.

我們完全致力於此。

796

01:21:12,000 --> 01:21:14,000

This is now Nemotron 3.

這就是現在的 Nemotron 3。

797

01:21:14,000 --> 01:21:17,000

We're currently working on Nemotron 4.

我們目前正在開發 Nemotron 4。

798

01:21:17,000 --> 01:21:34,000

So this entire toolkit from models, harnesses, tools and skills, and runtimes is the reason why every enterprise company in the world has the ability now to create their own agents.

所以這整個工具套件——模型、harness、工具與技能，以及 runtime——這就是為什麼全球每家企業公司現在都有能力創建自己的 Agent 的原因。

799

01:21:34,000 --> 01:21:37,000

Just like Cadence did with their super agents.

就像 Cadence 用他們的 SuperAgent 所做的那樣。

800

01:21:37,000 --> 01:21:39,000

And we're working with so many companies.

而且我們正在與如此多的公司合作。

801

01:21:39,000 --> 01:21:42,000

Cadence and CrowdStrike and Dessau and Palantir.

Cadence、CrowdStrike、Dassault 和 Palantir。

802

01:21:42,000 --> 01:21:44,000

SAP and ServiceNow.

SAP 和 ServiceNow。

803

01:21:44,000 --> 01:21:50,000

People always said, Jensen, the agents are going to disrupt these markets.

人們總是說：「Jensen，Agent 將顛覆這些市場。」

804

01:21:50,000 --> 01:21:54,000

I said completely the opposite and you can now see it.

我說完全相反，你們現在可以看到了。

805

01:21:54,000 --> 01:22:00,000

Agents is going to create the largest opportunity ever for my partners and friends.

Agent 將為我的合作夥伴和朋友創造有史以來最大的機會。

806

01:22:00,000 --> 01:22:08,000

And we have the NEMO, the NVIDIA agentic toolkit for enterprise AI to help them.

而我們有 NVIDIA 的企業 AI Agentic 工具套件來幫助他們。

807

01:22:08,000 --> 01:22:11,000

So there you go.

就是這樣。

808

01:22:15,000 --> 01:22:18,000

First, Vera Rubin in full production.

第一，Vera Rubin 全面量產。

809

01:22:18,000 --> 01:22:20,000

Two, Vera CPU.

第二，Vera CPU。

810

01:22:20,000 --> 01:22:24,000

CPU built for a new generation for agents.

為新一代 Agent 而建造的 CPU。

811

01:22:24,000 --> 01:22:35,000

And three, NVIDIA's enterprise AI toolkits so that every enterprise and every enterprise software company can build agents.

第三，NVIDIA 的企業 AI 工具套件，讓每家企業和每家企業軟體公司都能建構 Agent。

812

01:22:45,000 --> 01:22:50,000

My relationship with you started here.

我與你們的關係從這裡開始。

813

01:22:50,000 --> 01:22:58,000

And many of you, many of you, many of my friends and partners here in Taiwan, your companies started here.

你們許多人，你們許多人，我在台灣的許多朋友和合作夥伴，你們的公司也從這裡起步。

814

01:22:58,000 --> 01:23:06,000

This is in a lot of ways the beginning of the modern computer industry, 40 years now.

這在很多方面是現代電腦工業的開端，已經 40 年了。

815

01:23:06,000 --> 01:23:08,000

NVIDIA is 33 years old.

NVIDIA 已 33 歲了。

816

01:23:08,000 --> 01:23:17,000

The PC industry was already starting to get to Windows 1 and Windows 2 and Apple 1 and Apple 2.

當時 PC 行業已經開始進入 Windows 1、Windows 2、Apple 1 和 Apple 2 的時代。

817

01:23:17,000 --> 01:23:23,000

By the time that we came along, Windows 3.1 was the PC.

到我們出現時，Windows 3.1 就代表了 PC。

818

01:23:23,000 --> 01:23:28,000

And as you know, Windows 95 made PC personal.

如你所知，Windows 95 讓 PC 真正個人化。

819

01:23:28,000 --> 01:23:35,000

It took PC from enterprises, companies and made it into a consumer electronics device.

它將 PC 從企業和公司中拉出來，變成了一種消費電子設備。

820

01:23:35,000 --> 01:23:38,000

Everybody should have one and everybody does.

每個人都應該有一台，而且每個人都有了。

821

01:23:38,000 --> 01:23:40,000

This is the beginning.

這是開端。

822

01:23:40,000 --> 01:23:45,000

This computing platform did several things incredibly smart.

這個運算平台做了幾件非常聰明的事。

823

01:23:45,000 --> 01:23:48,000

Windows was not just disaggregated, as you know.

如你所知，Windows 不只是分解式的。

824

01:23:48,000 --> 01:23:51,000

Windows was properly abstracted.

Windows 有著恰當的抽象層次。

825

01:23:51,000 --> 01:23:53,000

It was architected just right.

它的架構設計恰到好處。

826

01:23:53,000 --> 01:24:02,000

Systems biases, open chip sets, the operating system with drivers,

系統偏好、開放晶片組、帶有驅動程式的作業系統，

827

01:24:02,000 --> 01:24:06,000

drivers that could be connected and installed at runtime,

可以在執行時連接和安裝的驅動程式，

828

01:24:06,000 --> 01:24:17,000

and an abstraction layer with a multimedia API that opened up the PC to what we all know today.

以及帶有多媒體 API 的抽象層，將 PC 開放成我們今天所知的樣子。

829

01:24:17,000 --> 01:24:23,000

Each one of these elements were essential in making the PC so popular.

這些元素中的每一個都是使 PC 如此普及的關鍵。

830

01:24:23,000 --> 01:24:31,000

40 years later, Microsoft and NVIDIA are going to reinvent the PC.

40 年後，Microsoft 和 NVIDIA 將重新發明 PC。

831

01:24:31,000 --> 01:24:34,000

This is going to be the new PC.

這將是新的 PC。

832

01:24:34,000 --> 01:24:38,000

Now tomorrow night, I think it's tomorrow night our time,

現在，明天晚上，我想是我們這裡的明天晚上，

833

01:24:38,000 --> 01:24:40,000

but I'm going to be with Satya.

我將和 Satya（微軟 CEO）在一起。

834

01:24:40,000 --> 01:24:44,000

We're going to talk a lot more about the work that we're doing together.

我們將更詳細地談談我們共同在做的工作。

835

01:24:44,000 --> 01:24:47,000

Microsoft and NVIDIA, over the last three years,

Microsoft 和 NVIDIA，在過去三年裡，

836

01:24:47,000 --> 01:24:55,000

it took this long to completely reinvent how the PC is going to work so that we could be ready for this moment.

花了這麼長時間完全重新發明 PC 的工作方式，這樣我們才能為這一刻做好準備。

837

01:24:55,000 --> 01:25:00,000

As I mentioned earlier, that compute pattern called the agent

正如我之前提到的，那種稱為 Agent 的運算模式

838

01:25:00,000 --> 01:25:03,000

is going to run in AI clouds.

將在 AI 雲端中運行。

839

01:25:03,000 --> 01:25:06,000

It's going to run inside enterprises.

它將在企業內部運行。

840

01:25:06,000 --> 01:25:10,000

It is also going to run on your PC.

它也將在你的 PC 上運行。

841

01:25:10,000 --> 01:25:14,000

What's going to happen to that PC when it has an autonomous agent,

當你的 PC 擁有一個自主 Agent 時，它將會發生什麼，

842

01:25:14,000 --> 01:25:18,000

an agent that's helping you, that understands you?

一個幫助你、理解你的 Agent？

843

01:25:18,000 --> 01:25:20,000

You could talk to it. It could look at you.

你可以和它說話。它可以看著你。

844

01:25:20,000 --> 01:25:27,000

You could ask it to read files, go help you, do some research.

你可以要求它讀取文件，去幫助你，做一些研究。

845

01:25:27,000 --> 01:25:29,000

It could do a lot more that I'll show you.

它還能做很多更多的事，我稍後會展示給你們看。

846

01:25:29,000 --> 01:25:37,000

But the new operating system is, of course, the old operating system plus large language models.

但新的作業系統當然是舊作業系統加上大型語言模型。

847

01:25:37,000 --> 01:25:43,000

Large language models in a lot of ways is the modern version of DirectX.

大型語言模型在很多方面是現代版的 DirectX。

848

01:25:43,000 --> 01:25:48,000

It has, of course, input and output, understands prompts, it understands computer vision,

它當然有輸入和輸出，理解 prompt，理解電腦視覺，

849

01:25:48,000 --> 01:25:51,000

it can generate video, it can generate sounds.

它可以生成影片，可以生成聲音。

850

01:25:51,000 --> 01:25:58,000

It is the modern extension, the intelligence extension of the PC, of a computer.

它是 PC、電腦的現代延伸，是智能的延伸。

851

01:25:58,000 --> 01:26:02,000

On top of that, the application, as I mentioned before,

在此之上，應用程式，如我之前所說，

852

01:26:02,000 --> 01:26:07,000

is going to be replaced by now an agentic runtime.

現在將被 Agentic runtime 所取代。

853

01:26:07,000 --> 01:26:10,000

And that is the modern application, an agent.

那就是現代應用程式——一個 Agent。

854

01:26:10,000 --> 01:26:13,000

Let's now take a look at what it can do.

讓我們來看看它能做什麼。

855

01:26:13,000 --> 01:26:19,000

It started with a spark, an idea,

一切從一個「spark」（靈感）開始，一個想法，

856

01:26:19,000 --> 01:26:23,000

to reimagine the PC for the first time in 40 years,

40 年來首次重新想像 PC，

857

01:26:23,000 --> 01:26:26,000

for the age of AI.

為了 AI 時代。

858

01:26:26,000 --> 01:26:31,000

What becomes of our personal computer in a world of agents?

在 Agent 的世界裡，我們的個人電腦將變成什麼？

859

01:26:31,000 --> 01:26:36,000

Agents running natively, connected to models, local or in the cloud,

Agent 本地原生運行，連接到本地或雲端的模型，

860

01:26:36,000 --> 01:26:43,000

our personal AI, sandboxed for security, running continuously, getting work done.

我們的個人 AI，以沙箱保護安全，持續運行，完成工作。

861

01:26:43,000 --> 01:26:47,000

The chips and the OS must evolve.

晶片和作業系統必須進化。

862

01:26:47,000 --> 01:26:55,000

Introducing RTX Spark, everything we've learned over 33 years, distilled into one chip.

推出 RTX Spark，我們 33 年來所學的一切，凝聚在一顆晶片中。

863

01:26:55,000 --> 01:27:04,000

Blackwell RTX GPU with 6,144 CUDA cores, one petaflop of AI performance.

Blackwell RTX GPU，搭載 6,144 個 CUDA 核心，一 petaflop 的 AI 性能。

864

01:27:04,000 --> 01:27:12,000

A custom 20-core Grace CPU, built in partnership with MediaTek, fused by MVLink.

定製的 20 核 Grace CPU，與聯發科技（MediaTek）合作建造，透過 NVLink 融合。

865

01:27:12,000 --> 01:27:17,000

128 gigabytes of unified memory.

128 GB 統一記憶體。

866

01:27:17,000 --> 01:27:20,000

TSMC 3-nanometer process.

台積電 3 奈米製程。

867

01:27:20,000 --> 01:27:23,000

70 billion transistors.

700 億個電晶體。

868

01:27:23,000 --> 01:27:30,000

And in close collaboration with Microsoft, a Windows platform for agents.

以及與 Microsoft 密切合作，一個用於 Agent 的 Windows 平台。

869

01:27:30,000 --> 01:27:34,000

We're reinventing the personal computer.

我們正在重新發明個人電腦。

870

01:27:34,000 --> 01:27:37,000

For creating.

為了創作。

871

01:27:37,000 --> 01:27:40,000

For gaming.

為了遊戲。

872

01:27:40,000 --> 01:27:43,000

For agents.

為了 Agent。

873

01:27:43,000 --> 01:27:47,000

This is the dawn of a new personal computing revolution.

這是新個人運算革命的黎明。

874

01:27:47,000 --> 01:27:51,000

And it starts with NVIDIA RTX Spark.

而它從 NVIDIA RTX Spark 開始。

875

01:28:05,000 --> 01:28:08,000

Here it is.

就是這個了。

876

01:28:08,000 --> 01:28:13,000

Of course, I got to show you the most beautiful part, which is video games.

當然，我要向你們展示最美麗的部分，那就是電子遊戲。

877

01:28:13,000 --> 01:28:16,000

It is also the closest to our heart.

這也是最貼近我們心靈的東西。

878

01:28:16,000 --> 01:28:18,000

This is Forza.

這是 Forza。

879

01:28:18,000 --> 01:28:21,000

This is 007, by the way.

對了，這是 007。

880

01:28:21,000 --> 01:28:24,000

The new 007 game, I'm looking forward to playing it.

新的 007 遊戲，我很期待玩它。

881

01:28:24,000 --> 01:28:27,000

I look a little bit like him.

我有點像他（笑）。

882

01:28:27,000 --> 01:28:32,000

Ladies and gentlemen, NVIDIA's RTX Spark laptops.

女士們先生們，NVIDIA 的 RTX Spark 筆記型電腦。

883

01:28:37,000 --> 01:28:43,000

Thank you.

謝謝。

884

01:28:43,000 --> 01:28:45,000

I have too many things in my pocket.

我口袋裡有太多東西了。

885

01:28:45,000 --> 01:28:48,000

Too many things.

太多東西了。

886

01:28:48,000 --> 01:28:49,000

OK.

好的。

887

01:28:49,000 --> 01:28:50,000

All right.

好的。

888

01:28:50,000 --> 01:28:53,000

This is the most amazing chip the world has ever built.

這是世界有史以來建造過的最令人驚嘆的晶片。

889

01:28:53,000 --> 01:28:59,000

This is the N1X that we built in partnership with MediaTek.

這是我們與聯發科技（MediaTek）合作建造的 N1X。

890

01:28:59,000 --> 01:29:01,000

I think I saw Rick earlier.

我想我之前看到 Rick（聯發科技代表）了。

891

01:29:01,000 --> 01:29:02,000

This is N1X.

這就是 N1X。

892

01:29:02,000 --> 01:29:03,000

This is a beautiful chip.

這是一顆美麗的晶片。

893

01:29:03,000 --> 01:29:10,000

This is a chip that, frankly, would take 33 years to build.

坦白說，這是一顆需要 33 年才能建造出來的晶片。

894

01:29:10,000 --> 01:29:16,000

And the reason for that is because 100% of NVIDIA software stack runs here.

原因是 100% 的 NVIDIA 軟體堆疊都可以在這裡運行。

895

01:29:16,000 --> 01:29:20,000

If you want to run digital biology, no problem.

如果你想運行數位生物學，沒問題。

896

01:29:20,000 --> 01:29:23,000

If you want to do seismic processing, no problem.

如果你想做地震處理，沒問題。

897

01:29:23,000 --> 01:29:25,000

You want astrophysics, no problem.

你想要天文物理，沒問題。

898

01:29:25,000 --> 01:29:31,000

Everything associated with CUDA, all the physics, all the biology, all the genomics, all the AI, no problem.

所有與 CUDA 相關的東西，所有物理、生物、基因體學、AI，沒問題。

899

01:29:31,000 --> 01:29:33,000

All the computer graphics, no problem.

所有電腦圖形，沒問題。

900

01:29:33,000 --> 01:29:43,000

Every single application NVIDIA has ever created and every single application that Windows has ever run,

NVIDIA 有史以來創建的每一個應用程式，以及 Windows 有史以來運行的每一個應用程式，

901

01:29:43,000 --> 01:29:54,000

Microsoft and NVIDIA meticulously optimized everything so that this computer literally runs everything the world has ever created.

Microsoft 和 NVIDIA 精心優化了一切，使這台電腦真正能夠運行世界有史以來創造的一切。

902

01:29:54,000 --> 01:29:59,000

Plus, it now runs Agents, an incredible computer.

加上，它現在還可以運行 Agent，一台令人難以置信的電腦。

903

01:29:59,000 --> 01:30:01,000

I'm so proud of it.

我為此感到非常自豪。

904

01:30:07,000 --> 01:30:08,000

Okay.

好的。

905

01:30:08,000 --> 01:30:14,000

Now, I want you to keep that in mind in the next video I'm going to show you.

現在，我希望你在接下來我要展示的影片中記住這一點。

906

01:30:14,000 --> 01:30:17,000

Just imagine everything here is going to run on your PC.

想像一下，這裡的一切都將在你的 PC 上運行。

907

01:30:17,000 --> 01:30:25,000

Now, that computer could have a local Nemotron 3 Ultra model or Nemotron 3 Super model,

現在，那台電腦可以有一個本地的 Nemotron 3 Ultra 或 Nemotron 3 Super 模型，

908

01:30:25,000 --> 01:30:33,000

or it could have a Cloud Code or Codex or some other model in the Cloud or something on the network.

或者它可以有 Claude Code 或 Codex 或雲端中的其他模型，或網路上的某些東西。

909

01:30:33,000 --> 01:30:36,000

And it's going to work and do something amazing.

它將工作並做出令人驚嘆的事情。

910

01:30:36,000 --> 01:30:37,000

Let's play.

讓我們播放吧。

911

01:30:37,000 --> 01:30:39,000

[Video playing]

【影片播放中】

912

01:30:39,000 --> 01:30:42,000

Every house starts as an idea.

每棟房子都從一個想法開始。

913

01:30:42,000 --> 01:30:51,000

Getting from idea to design takes a myriad of tools, expertise, and a lot of time.

從想法到設計需要無數工具、專業知識和大量時間。

914

01:30:51,000 --> 01:30:58,000

Now, an agent running locally on RTX Spark can help me design a house using the tools on my laptop

現在，在 RTX Spark 上本地運行的 Agent 可以幫我使用筆記型電腦上的工具設計一棟房子，

915

01:30:58,000 --> 01:31:04,000

with an open shell sandbox running the Hermes harness connected to CloudSonic in the Cloud.

搭配 Open Shell 沙箱，運行連接到雲端 Claude Sonnet 的 Hermes harness。

916

01:31:04,000 --> 01:31:11,000

I select the site, share my concept sketches and mood board of styles to inspire my design,

我選擇地點，分享我的概念草圖和風格情緒板來啟發我的設計，

917

01:31:11,000 --> 01:31:19,000

and the prompt, a text description of the requirements and the design intent.

以及 prompt，需求和設計意圖的文字描述。

918

01:31:19,000 --> 01:31:22,000

My agent goes to work using the tools on my laptop.

我的 Agent 開始使用筆記型電腦上的工具工作。

919

01:31:22,000 --> 01:31:30,000

It opens Rhino and starts modeling the site, shaping terrain, setbacks, and the building envelope.

它打開 Rhino 開始建模基地，塑造地形、退縮線和建築外殼。

920

01:31:30,000 --> 01:31:37,000

Then it proposes building forms optimized for cost, comfort, and quality.

然後它提出針對成本、舒適度和品質進行優化的建築形式。

921

01:31:37,000 --> 01:31:43,000

With the form defined, my agent generates the interior layout, walls, circulation.

形式確定後，我的 Agent 生成室內佈局、牆壁、動線。

922

01:31:43,000 --> 01:31:45,000

Rooms begin to take shape.

房間開始成形。

923

01:31:45,000 --> 01:31:52,000

I jump in whenever I want to adjust to change.

我可以隨時跳進來進行調整和修改。

924

01:31:52,000 --> 01:31:56,000

Doors, windows, and structural elements are placed automatically.

門窗和結構元素自動放置。

925

01:31:56,000 --> 01:32:03,000

My agent detects its own mistakes and fixes them.

我的 Agent 自我檢測錯誤並修復它們。

926

01:32:03,000 --> 01:32:08,000

When I approve, the agent exports the model from Rhino into Blender.

當我批准後，Agent 將模型從 Rhino 導出到 Blender。

927

01:32:08,000 --> 01:32:13,000

Materials and object properties transfer with the design context intact.

材質和物件屬性在設計上下文完整的情況下傳遞。

928

01:32:13,000 --> 01:32:17,000

I fine-tune the materials, get the look just right.

我微調材質，讓外觀恰到好處。

929

01:32:17,000 --> 01:32:19,000

Then I pick the shots.

然後我選擇拍攝角度。

930

01:32:19,000 --> 01:32:21,000

Blender renders the house.

Blender 渲染房子。

931

01:32:21,000 --> 01:32:26,000

My agent, using generative AI with the Flux 2 model, makes them photoreal.

我的 Agent 使用 Flux 2 生成式 AI 模型，使它們達到照片級真實感。

932

01:32:26,000 --> 01:32:29,000

Multiple viewpoints, lighting conditions.

多個視角，多種光照條件。

933

01:32:29,000 --> 01:32:34,000

What was once a complex workflow is now guided and simplified by my agent,

過去複雜的工作流程現在由我的 Agent 引導和簡化，

934

01:32:34,000 --> 01:32:41,000

working with an RTX Spark, designed at the speed of imagination.

配合 RTX Spark，以想像力的速度進行設計。

935

01:32:47,000 --> 01:32:49,000

He's seeing the world of agents.

他正在見證 Agent 的世界。

936

01:32:49,000 --> 01:32:51,000

The developers are so excited about it.

開發者們對此非常興奮。

937

01:32:51,000 --> 01:32:53,000

This is an incredible computer.

這是一台令人難以置信的電腦。

938

01:32:53,000 --> 01:32:57,000

All of the acceleration, all the software capabilities associated with it,

所有的加速，所有與它相關的軟體能力，

939

01:32:57,000 --> 01:33:02,000

working with every developer to make it incredible for all of you.

與每一位開發者合作，使它對你們所有人都令人難以置信。

940

01:33:02,000 --> 01:33:05,000

The next one, Adobe.

下一個，Adobe。

941

01:33:05,000 --> 01:33:09,000

Incredible tool suite, of course, used by tens of millions of people around the world.

令人難以置信的工具套件，當然，被全球數千萬人使用。

942

01:33:09,000 --> 01:33:15,000

They have re-engineered the architecture, the core of Adobe Photoshop and Premiere,

他們重新設計了架構，Adobe Photoshop 和 Premiere 的核心，

943

01:33:15,000 --> 01:33:18,000

and they're going to release it for RTX Spark.

他們將為 RTX Spark 發布它。

944

01:33:18,000 --> 01:33:20,000

It is twice as fast. It's already fast.

它快了兩倍。它本來就已經很快了。

945

01:33:20,000 --> 01:33:22,000

Now it's going to be twice as fast.

現在它將快兩倍。

946

01:33:22,000 --> 01:33:26,000

And it's also designed to be agent-friendly.

而且它也被設計為對 Agent 友好的。

947

01:33:26,000 --> 01:33:31,000

With its MCP server, it can now interact with agents on your laptop.

通過其 MCP 伺服器，它現在可以與你筆記型電腦上的 Agent 互動。

948

01:33:31,000 --> 01:33:38,000

The number of customers, the number of partners that are so excited to bring RTX Spark to the market

如此多的客戶、如此多的合作夥伴都迫不及待地要將 RTX Spark 推向市場，

949

01:33:38,000 --> 01:33:40,000

is just incredible.

真的令人難以置信。

950

01:33:40,000 --> 01:33:48,000

You know, this is the first across the lineup of PC reinvention for 40 years.

你知道，這是 40 年來 PC 全線重新發明的第一次。

951

01:33:48,000 --> 01:33:54,000

And I'm just so happy that all of you and the ecosystem around the world has joined us.

我很高興你們所有人和全球生態系統都加入了我們。

952

01:33:54,000 --> 01:33:56,000

This is basically everybody.

基本上是所有人。

953

01:33:56,000 --> 01:34:02,000

Everybody will support RTX Spark and will be building incredibly smart and powerful

每個人都將支持 RTX Spark，並將建造極其智能和強大的

954

01:34:02,000 --> 01:34:04,000

and beautiful laptops with all of us.

與我們一起打造美麗的筆記型電腦。

955

01:34:04,000 --> 01:34:06,000

Thank you very much.

非常感謝。

956

01:34:11,000 --> 01:34:15,000

But that's not all. That's not all.

但這還不是全部。這還不是全部。

957

01:34:15,000 --> 01:34:19,000

RTX Spark is a reinvention of laptop.

RTX Spark 是筆記型電腦的重新發明。

958

01:34:19,000 --> 01:34:24,000

But in fact, Microsoft NVIDIA is reinventing all of PC.

但事實上，Microsoft 和 NVIDIA 正在重新發明整個 PC。

959

01:34:24,000 --> 01:34:28,000

And today we're announcing a whole new line.

今天我們宣布一整個全新產品線。

960

01:34:28,000 --> 01:34:31,000

Three revolutionary Windows machines

三款革命性的 Windows 機器，

961

01:34:31,000 --> 01:34:36,000

covering desktop, laptop, and workstations.

涵蓋桌機、筆電和工作站。

962

01:34:36,000 --> 01:34:43,000

All 100% Windows compatible, 100% CUDA, 100% NVIDIA AI Tensor Core.

全部 100% Windows 相容、100% CUDA、100% NVIDIA AI Tensor Core。

963

01:34:43,000 --> 01:34:47,000

Everything that runs, that you see that runs on NVIDIA

所有在 NVIDIA 上運行的一切，你所見到在 NVIDIA 上運行的一切，

964

01:34:47,000 --> 01:34:51,000

and all these different platforms around the world runs here.

以及全球所有這些不同平台上的一切，都可以在這裡運行。

965

01:34:51,000 --> 01:34:58,000

This is the first completely re-engineered, reinvented line of PCs

這是第一條完全重新設計、重新發明的 PC 產品線，

966

01:34:58,000 --> 01:35:00,000

that has happened in 40 years.

這在 40 年來是第一次發生。

967

01:35:00,000 --> 01:35:02,000

Now what's really amazing is this.

真正令人驚嘆的是這個。

968

01:35:02,000 --> 01:35:05,000

So this is the RTX Spark laptop.

這是 RTX Spark 筆記型電腦。

969

01:35:05,000 --> 01:35:09,000

This is the desktop.

這是桌機。

970

01:35:09,000 --> 01:35:11,000

So this one's from MSI.

這一台是 MSI 的。

971

01:35:11,000 --> 01:35:13,000

Joseph, this one's yours.

Joseph，這台是你的。

972

01:35:13,000 --> 01:35:16,000

Okay. Look how beautiful it is.

好的。看看它多美麗。

973

01:35:16,000 --> 01:35:19,000

This agent could run 24/7.

這個 Agent 可以 24/7 全天候運行。

974

01:35:19,000 --> 01:35:21,000

Meter free.

無計量（不按時間計費）。

975

01:35:21,000 --> 01:35:23,000

And you could download your agent.

你可以下載你的 Agent。

976

01:35:23,000 --> 01:35:28,000

You could raise your lobster in here.

你可以在這裡「養你的 AI 龍蝦」（raise your lobster：幽默比喻，指培養/養成你的 AI 代理）。

977

01:35:28,000 --> 01:35:31,000

This is your clock. It's running all the time.

這是你的「時鐘」。它一直在運行。

978

01:35:31,000 --> 01:35:34,000

No meter anxiety.

沒有計量焦慮（不用擔心按量計費）。

979

01:35:34,000 --> 01:35:37,000

And it's sitting here connected to your whole house,

它坐在這裡，連接到你的整個家，

980

01:35:37,000 --> 01:35:41,000

connected to your laptop, connected to your display,

連接到你的筆記型電腦、連接到你的螢幕，

981

01:35:41,000 --> 01:35:47,000

all the cameras, your dryer, your water cooler, your water heater,

所有攝影機、你的烘乾機、飲水機、熱水器，

982

01:35:47,000 --> 01:35:49,000

your everything, whatever you want.

你的一切，隨你所願。

983

01:35:49,000 --> 01:35:51,000

Your security system, all connected to this.

你的安全系統，全都連接到這個。

984

01:35:51,000 --> 01:35:56,000

And this becomes your personal AI, your personal AI agent.

這就成為你的個人 AI，你的個人 AI Agent。

985

01:35:56,000 --> 01:35:59,000

And it gets smarter and smarter and smarter over time

它隨著時間越來越聰明，

986

01:35:59,000 --> 01:36:02,000

because today we have Neemotron 3 Ultra.

因為今天我們有 Nemotron 3 Ultra。

987

01:36:02,000 --> 01:36:04,000

Tomorrow we have Neemotron 4.

明天我們有 Nemotron 4。

988

01:36:04,000 --> 01:36:06,000

And then Neemotron 5, Neemotron 6.

然後是 Nemotron 5、Nemotron 6。

989

01:36:06,000 --> 01:36:09,000

And we just keep getting smarter and smarter and smarter.

我們就一直越來越聰明、越來越聰明。

990

01:36:09,000 --> 01:36:12,000

And meanwhile this is sitting at home helping you do things.

同時這台機器坐在家裡幫你做事情。

991

01:36:12,000 --> 01:36:14,000

If you want a book of travel, no problem.

如果你想預訂旅遊行程，沒問題。

992

01:36:14,000 --> 01:36:21,000

And if you want an incredible system,

如果你想要一個令人難以置信的系統，

993

01:36:21,000 --> 01:36:27,000

this is a DGX station for Windows, compatible with Windows,

這是 DGX Station for Windows，與 Windows 相容，

994

01:36:27,000 --> 01:36:35,000

runs everything in Windows, and it has 768 gigabytes of memory.

在 Windows 上運行一切，它有 768 GB 記憶體。

995

01:36:35,000 --> 01:36:39,000

And so you could run a trillion parameter model.

所以你可以運行一兆參數的模型。

996

01:36:39,000 --> 01:36:41,000

This is unbelievable.

這令人難以置信。

997

01:36:41,000 --> 01:36:46,000

20 petaflops, 8 terabytes per second of memory bandwidth.

20 petaflops，每秒 8 TB 的記憶體頻寬。

998

01:36:46,000 --> 01:36:49,000

And this sits by your desk.

它就坐在你的桌旁。

999

01:36:49,000 --> 01:36:52,000

You basically, if you're a developer of large language models,

基本上，如果你是大型語言模型的開發者，

1000

01:36:52,000 --> 01:36:57,000

you're a developer of agents, having this sit by your desk

你是 Agent 的開發者，讓這台機器坐在你的桌旁，

1001

01:36:57,000 --> 01:36:59,000

gives you all the compute you need.

給你所需的所有算力。

1002

01:36:59,000 --> 01:37:02,000

And then when you deploy it, you put it into the cloud.

然後當你部署時，把它放到雲端。

1003

01:37:02,000 --> 01:37:06,000

Now there's something that if you look at this and think about this,

現在有一件事，如果你看著這個並思考，

1004

01:37:06,000 --> 01:37:09,000

something is happening here.

這裡正在發生某些事情。

1005

01:37:09,000 --> 01:37:17,000

Remember, 15, 20 years ago, we used to have an idea called a foam.

記住，15、20 年前，我們有一個叫做「手機」（phone）的概念。

1006

01:37:17,000 --> 01:37:21,000

Today we have an idea called a PC.

今天我們有一個叫做 PC 的東西。

1007

01:37:21,000 --> 01:37:24,000

Today when you think about your phone,

今天當你想到你的手機時，

1008

01:37:24,000 --> 01:37:31,000

the one thing you don't do with it is make phone calls.

你幾乎不用它做的那一件事是打電話。

1009

01:37:31,000 --> 01:37:33,000

You do just about everything else.

你用它做幾乎所有其他的事。

1010

01:37:33,000 --> 01:37:40,000

And so that phone means something very different to you than a phone of the past.

所以那部手機對你來說，與過去的電話有著截然不同的意義。

1011

01:37:40,000 --> 01:37:45,000

I am certain what's going to happen here is that the PC 10 years from now

我確信這裡即將發生的是，10 年後的 PC

1012

01:37:45,000 --> 01:37:48,000

and the PC that you think about today, a tool,

與你今天所認識的 PC——一個工具，

1013

01:37:48,000 --> 01:37:54,000

where you launch applications, click and type,

你在其中啟動應用程式、點擊和輸入——

1014

01:37:54,000 --> 01:37:58,000

and this PC is going to be completely different.

那台 PC 將會是完全不同的東西。

1015

01:37:58,000 --> 01:38:00,000

Here's my theory.

這是我的理論。

1016

01:38:00,000 --> 01:38:06,000

I can totally imagine just as every house today has a home theater,

我完全可以想像，就像今天每個家庭都有家庭劇院，

1017

01:38:06,000 --> 01:38:14,000

where many houses have home theaters, big TVs, lawn mowers, dishwashers.

許多房屋有家庭劇院、大螢幕電視、割草機、洗碗機。

1018

01:38:14,000 --> 01:38:20,000

I could totally imagine that someday there's actually an AI supercomputer in your house.

我完全可以想像，有一天你家裡實際上有一台 AI 超級電腦。

1019

01:38:20,000 --> 01:38:22,000

And it's running all of your agents.

它正在運行你所有的 Agent。

1020

01:38:22,000 --> 01:38:24,000

It's running all of your assistants.

它正在運行你所有的助理。

1021

01:38:24,000 --> 01:38:28,000

And they're doing all kinds of things for you all the time.

而且它們一直在為你做各種各樣的事情。

1022

01:38:28,000 --> 01:38:32,000

And you have to have it in your house just like you have a home theater in your house.

你必須在家裡擁有它，就像你家裡有家庭劇院一樣。

1023

01:38:32,000 --> 01:38:36,000

You have stereos in your house. You have game consoles in your house.

你家裡有音響。你家裡有遊戲主機。

1024

01:38:36,000 --> 01:38:40,000

You would assist AI agent computers running in your house.

你家裡會有 AI Agent 電腦在運行。

1025

01:38:40,000 --> 01:38:47,000

And these, in time, becomes a lot more like R2D2 to you.

而這些，隨著時間推移，對你來說越來越像 R2-D2。

1026

01:38:47,000 --> 01:38:54,000

It becomes more like C3PO to you than it feels like a PC to you.

它對你來說越來越像 C-3PO，而不是感覺像一台 PC。

1027

01:38:54,000 --> 01:38:59,000

There is no question this reinvention of the computer is as big of a deal

毫無疑問，這次電腦的重新發明是如此重大的事，

1028

01:38:59,000 --> 01:39:04,000

as the reinvention of the phone into what we now know as the smartphone.

就像把電話重新發明成我們現在所知道的智慧型手機一樣重要。

1029

01:39:04,000 --> 01:39:06,000

And so this is the beginning of that journey.

所以這是那段旅程的開端。

1030

01:39:06,000 --> 01:39:09,000

This is the beginning of a new line.

這是新產品線的開端。

1031

01:39:09,000 --> 01:39:11,000

And so we have a roadmap for this.

我們有這方面的路線圖。

1032

01:39:11,000 --> 01:39:14,000

This is a brand new product family for us.

這是我們全新的產品系列。

1033

01:39:14,000 --> 01:39:21,000

Every single generation of architecture, we will have a desktop, a laptop, a workstation,

每一個架構世代，我們都將有桌機、筆電、工作站，

1034

01:39:21,000 --> 01:39:25,000

and then a desktop, a laptop, and workstation.

然後是桌機、筆電和工作站。

1035

01:39:25,000 --> 01:39:29,000

And the thing that I am just incredibly pleased, incredibly honored,

而讓我感到無比高興、無比榮幸的是，

1036

01:39:29,000 --> 01:39:36,000

is that 100% of the world's PC industry has joined us to reinvent the PC.

全球 100% 的 PC 行業都加入了我們重新發明 PC 的行列。

1037

01:39:36,000 --> 01:39:38,000

A new line, a new beginning.

新的產品線，新的開始。

1038

01:39:38,000 --> 01:39:40,000

Thank you.

謝謝。

1039

01:39:47,000 --> 01:39:59,000

As you know, agentic AI is just a digital robot.

如你所知，Agentic AI 就是一個數位機器人。

1040

01:39:59,000 --> 01:40:06,000

It understands, it reasons, it plans, and it acts and uses tools.

它理解、推理、規劃，並採取行動和使用工具。

1041

01:40:06,000 --> 01:40:11,000

Agentic AI is going to run across all of these computers.

Agentic AI 將在所有這些電腦上運行。

1042

01:40:11,000 --> 01:40:15,000

And you've seen me talk about each and every one of these over time.

你們看到我一次又一次地談論每一個這樣的東西。

1043

01:40:15,000 --> 01:40:20,000

We're working on human or robotics computers, robotics computers of all kinds.

我們正在研究人形或機器人電腦，各種各樣的機器人電腦。

1044

01:40:20,000 --> 01:40:22,000

We're working on self-driving car computers.

我們正在研究自動駕駛汽車電腦。

1045

01:40:22,000 --> 01:40:24,000

We're working on satellites.

我們正在研究衛星。

1046

01:40:24,000 --> 01:40:27,000

You have GeForce, which has TensorCores.

你有 GeForce，它有 TensorCore。

1047

01:40:27,000 --> 01:40:30,000

I just talked about a whole new line of PCs.

我剛剛談到了一整條全新的 PC 產品線。

1048

01:40:30,000 --> 01:40:34,000

Agriculture equipment, manufacturing equipment, heavy industry equipment

農業設備、製造設備、重工業設備，

1049

01:40:34,000 --> 01:40:36,000

will all be agentic.

全都將是 Agentic 的。

1050

01:40:36,000 --> 01:40:41,000

You'll even have a little agentic helper for yourself.

你甚至會有一個小小的 Agentic 助理為你服務。

1051

01:40:41,000 --> 01:40:47,000

Even your base stations, the radio stations of the future, are going to be agentic.

甚至你的基地台，未來的無線電基站，都將是 Agentic 的。

1052

01:40:47,000 --> 01:40:53,000

Understanding traffic and thinking about how to coordinate with the other base stations

理解流量並思考如何與其他基站協調，

1053

01:40:53,000 --> 01:40:58,000

so that you could use as little energy as possible, increase the utilization,

這樣你就可以使用盡可能少的能量，提高利用率，

1054

01:40:58,000 --> 01:41:01,000

the efficiency of the spectral efficiency.

提升頻譜效率。

1055

01:41:01,000 --> 01:41:04,000

And so everything will run agents.

所以一切都將運行 Agent。

1056

01:41:04,000 --> 01:41:07,000

Today, NVIDIA is largely in the center.

今天，NVIDIA 在很大程度上處於中心位置。

1057

01:41:07,000 --> 01:41:15,000

But I am pretty certain that there will be tens of billions, hundreds of billions over time

但我相當確定，隨著時間推移，將會有數百億、數千億

1058

01:41:15,000 --> 01:41:20,000

of agentic systems, agentic computers that are going to be running around the world.

個 Agentic 系統、Agentic 電腦在全球各地運行。

1059

01:41:20,000 --> 01:41:23,000

The biggest problem is data.

最大的問題是資料。

1060

01:41:23,000 --> 01:41:28,000

In the case of language models, all the English and all the language that we have on the internet

以語言模型為例，我們在網路上擁有的所有英語和所有語言，

1061

01:41:28,000 --> 01:41:31,000

that we trained on was from the perspective of us.

我們訓練用的資料都是從我們人類的視角出發的。

1062

01:41:31,000 --> 01:41:33,000

We wrote it and we're reading it.

我們寫了它，我們在閱讀它。

1063

01:41:33,000 --> 01:41:41,000

However, in order to create data for AI robotics, it has to be in the perception,

然而，為了為 AI 機器人創建資料，必須從感知的視角出發，

1064

01:41:41,000 --> 01:41:43,000

the perspective of the robot.

即機器人的視角。

1065

01:41:43,000 --> 01:41:48,000

And most of the world's video data is from a third person, not first person.

而世界上大多數影片資料是第三人稱視角，而不是第一人稱。

1066

01:41:48,000 --> 01:41:57,000

And so agentic systems, robotic systems, physical AI, the data is the hardest problem.

所以對於 Agentic 系統、機器人系統、Physical AI 來說，資料是最難解決的問題。

1067

01:41:57,000 --> 01:42:00,000

You've seen us move up this ladder.

你們看到我們沿著這個階梯向上爬。

1068

01:42:00,000 --> 01:42:04,000

We started with teleoperations, which is basically human demonstration.

我們從遠程操作開始，基本上就是人類示範。

1069

01:42:04,000 --> 01:42:09,000

This is no different than the big breakthrough of reinforcement learning human feedback.

這與強化學習人類反饋（RLHF）的重大突破沒有不同。

1070

01:42:09,000 --> 01:42:12,000

This, then we use simulation.

然後我們使用模擬。

1071

01:42:12,000 --> 01:42:14,000

This is where Omniverse comes in.

這就是 Omniverse 發揮作用的地方。

1072

01:42:14,000 --> 01:42:19,000

This is no different than reinforcement learning verifiable rewards.

這與強化學習可驗證獎勵（RLVR）沒有不同。

1073

01:42:19,000 --> 01:42:27,000

And so we use these systems to bootstrap the AI model, the physical AI model.

所以我們使用這些系統來引導 AI 模型，即 Physical AI 模型。

1074

01:42:27,000 --> 01:42:34,000

Eventually, we're able to learn from third person, reprojecting it into first person.

最終，我們能夠從第三人稱學習，將其重新投影為第一人稱。

1075

01:42:34,000 --> 01:42:40,000

And now, eventually, through bootstrapping, we have a world foundation model

而現在，最終通過引導，我們有了一個世界基礎模型，

1076

01:42:40,000 --> 01:42:45,000

that can understand the physical world from any perspective you want.

它可以從任何你想要的視角理解物理世界。

1077

01:42:45,000 --> 01:42:50,000

Third person, first person, outside in, inside out, doesn't matter.

第三人稱、第一人稱、由外到內、由內到外，都沒關係。

1078

01:42:50,000 --> 01:42:53,000

This is a big breakthrough indeed.

這確實是一個重大突破。

1079

01:42:53,000 --> 01:42:59,000

And today, we are announcing Cosmos 3.

而今天，我們宣布 Cosmos 3。

1080

01:42:59,000 --> 01:43:04,000

Cosmos 3 is the frontier of physical AI.

Cosmos 3 是 Physical AI 的前沿。

1081

01:43:04,000 --> 01:43:08,000

We are at the frontier with language models.

在語言模型方面，我們處於前沿。

1082

01:43:08,000 --> 01:43:10,000

There are so many people working on it.

有如此多的人在研究它。

1083

01:43:10,000 --> 01:43:14,000

However, in physical AI, we are absolutely the world's best.

然而，在 Physical AI 方面，我們絕對是世界最佳。

1084

01:43:14,000 --> 01:43:16,000

I am so proud of the team for doing this.

我為團隊做到這一點感到非常自豪。

1085

01:43:16,000 --> 01:43:19,000

This is the foundation model for all of your work.

這是你們所有工作的基礎模型。

1086

01:43:19,000 --> 01:43:24,000

Whenever you want to create a robot, whenever you want to create a factory robot

每當你想創建一個機器人，每當你想創建一個工廠機器人，

1087

01:43:24,000 --> 01:43:31,000

or a robot that works in a factory, any kind of robot that involves physical world,

或在工廠工作的機器人，任何涉及物理世界的機器人，

1088

01:43:31,000 --> 01:43:36,000

you now have a companion, a Cosmos 3 that can understand and reason,

你現在有了一個伴侶，一個可以理解和推理的 Cosmos 3，

1089

01:43:36,000 --> 01:43:43,000

it can generate, it can simulate in the loop, it can even be the policy itself.

它可以生成、可以在循環中模擬，甚至可以成為策略本身。

1090

01:43:43,000 --> 01:43:46,000

It is on the top of leaderboards all over the world.

它在全球各地的排行榜上名列前茅。

1091

01:43:46,000 --> 01:43:48,000

I am incredibly proud of Cosmos.

我為 Cosmos 感到無比自豪。

1092

01:43:48,000 --> 01:43:52,000

And today, we are announcing Cosmos 3. Let's take a look.

而今天，我們宣布 Cosmos 3。讓我們來看看。

1093

01:43:52,000 --> 01:43:56,000

The real world is infinite and unpredictable.

真實世界是無限且不可預測的。

1094

01:43:56,000 --> 01:44:02,000

Physical AI needs data, but real world data is impossible to scale.

Physical AI 需要資料，但真實世界的資料無法規模化。

1095

01:44:02,000 --> 01:44:06,000

For physical AI, compute is data.

對於 Physical AI 來說，算力就是資料。

1096

01:44:06,000 --> 01:44:11,000

This is Cosmos, an open frontier omni-model for physical AI.

這就是 Cosmos，一個面向 Physical AI 的開放前沿全能模型。

1097

01:44:11,000 --> 01:44:14,000

Built on a new mixture of transformers architecture,

建立在新的混合 Transformer 架構上，

1098

01:44:14,000 --> 01:44:19,000

pixels, action, sound and language flow into the autoregressive transformer,

像素、動作、聲音和語言流入自回歸 Transformer，

1099

01:44:19,000 --> 01:44:23,000

which reasons, plans and instructs the diffusion transformer,

它進行推理、規劃並指導擴散 Transformer，

1100

01:44:23,000 --> 01:44:26,000

which generates what comes next.

後者生成接下來發生的事情。

1101

01:44:26,000 --> 01:44:31,000

Developers post-train Cosmos across embodiments and use cases.

開發者針對不同具身形態和使用案例對 Cosmos 進行後訓練。

1102

01:44:31,000 --> 01:44:37,000

As a VLM, Cosmos watches the physical world, understands what's happening,

作為 VLM（視覺語言模型），Cosmos 觀察物理世界，理解正在發生的事情，

1103

01:44:37,000 --> 01:44:41,000

describing scenes and flagging what matters.

描述場景並標記重要事項。

1104

01:44:41,000 --> 01:44:46,000

As a world model, Cosmos generates physics-accurate synthetic video

作為世界模型，Cosmos 生成物理精確的合成影片，

1105

01:44:46,000 --> 01:44:50,000

from an image, text or video.

從圖像、文字或影片生成。

1106

01:44:50,000 --> 01:44:55,000

As a simulator, Cosmos closes the loop for policy training and evaluation.

作為模擬器，Cosmos 為策略訓練和評估形成閉環。

1107

01:44:55,000 --> 01:45:00,000

And as the foundation of NVIDIA Omnidreams, an action-conditioned world model,

作為 NVIDIA Omnidreams 的基礎，一個動作條件化的世界模型，

1108

01:45:00,000 --> 01:45:05,000

Cosmos predicts the future frame by frame.

Cosmos 逐幀預測未來。

1109

01:45:05,000 --> 01:45:09,000

Post-train Cosmos and it becomes a world action model.

對 Cosmos 進行後訓練，它就成為一個世界動作模型。

1110

01:45:09,000 --> 01:45:15,000

Perceiving, reasoning, planning, generating actions.

感知、推理、規劃、生成動作。

1111

01:45:15,000 --> 01:45:18,000

For robots of every kind.

適用於各種機器人。

1112

01:45:18,000 --> 01:45:22,000

For everything that moves.

適用於所有會移動的事物。

1113

01:45:22,000 --> 01:45:29,000

A new kind of data, a new kind of teacher, generated by compute.

一種新型資料，一種新型教師，由算力生成。

1114

01:45:29,000 --> 01:45:35,000

Cosmos. The foundation for developers of the age of physical AI.

Cosmos。Physical AI 時代開發者的基礎。

1115

01:45:35,000 --> 01:45:52,000

It takes data plus compute, gives you AI.

用資料加算力，給你 AI。

1116

01:45:52,000 --> 01:45:56,000

Now that we have AI, compute is data.

現在我們有了 AI，算力就是資料。

1117

01:45:56,000 --> 01:46:00,000

And so, use Cosmos 3, train a whole bunch of AI models.

所以，使用 Cosmos 3，訓練大量 AI 模型。

1118

01:46:00,000 --> 01:46:04,000

Cosmos is such an incredible open model system, it's exactly the same as Neymotron.

Cosmos 是如此令人難以置信的開放模型系統，完全和 Nemotron 一樣。

1119

01:46:04,000 --> 01:46:09,000

We open the model, we open the data, and we even open how we trained it

我們開放模型，我們開放資料，我們甚至開放我們的訓練方式，

1120

01:46:09,000 --> 01:46:14,000

so that you could enhance it for yourself and turn Cosmos into your proprietary model.

這樣你就可以自己增強它，把 Cosmos 變成你的專有模型。

1121

01:46:14,000 --> 01:46:19,000

We have such incredible partners working with us in so many different industries.

我們在如此多不同行業都有令人難以置信的合作夥伴與我們合作。

1122

01:46:19,000 --> 01:46:25,000

Now the model itself is, of course, the most understandable part of the AI stack.

當然，模型本身是 AI 堆疊中最容易理解的部分。

1123

01:46:25,000 --> 01:46:27,000

But the AI stack is very complicated.

但 AI 堆疊非常複雜。

1124

01:46:27,000 --> 01:46:34,000

It has generators, the model, simulators, and the runtime.

它有生成器、模型、模擬器和 runtime。

1125

01:46:34,000 --> 01:46:41,000

Just as it is for agentic systems, these cars are essentially a physical AI,

就像 Agentic 系統一樣，這些汽車本質上是一個 Physical AI，

1126

01:46:41,000 --> 01:46:48,000

a agentic robot that is an autonomous vehicle, has also this complicated stack.

一個 Agentic 機器人，即自動駕駛車輛，也有這個複雜的堆疊。

1127

01:46:48,000 --> 01:46:55,000

Today we're announcing AlpaMile 2, an open model for self-driving cars.

今天我們宣布 AlpaMile 2，一個用於自動駕駛汽車的開放模型。

1128

01:46:55,000 --> 01:46:59,000

We're working with car companies across the world.

我們正在與全球各地的汽車公司合作。

1129

01:46:59,000 --> 01:47:02,000

If you look at these brands that have signed up for the NVIDIA Hyperion,

如果你看看這些簽約 NVIDIA Hyperion 的品牌，

1130

01:47:02,000 --> 01:47:11,000

that are building NVIDIA Hyperion cars, this represents about 80% of the world's cars.

正在建造 NVIDIA Hyperion 汽車，這代表了全球約 80% 的汽車。

1131

01:47:11,000 --> 01:47:14,000

The manufacturers represent 80% of the world's cars.

這些製造商代表了全球 80% 的汽車。

1132

01:47:14,000 --> 01:47:20,000

We are going to have a whole lot of NVIDIA Hyperion systems that are able to run AlpaMile

我們將擁有大量能夠運行 AlpaMile 的 NVIDIA Hyperion 系統，

1133

01:47:20,000 --> 01:47:23,000

or anybody else's AV stack.

或任何其他人的自動駕駛（AV）堆疊。

1134

01:47:23,000 --> 01:47:26,000

We are also connected into mobility services.

我們也連接到了移動服務。

1135

01:47:26,000 --> 01:47:31,000

Approximately 97% of the world's mobility services are connecting with us

全球約 97% 的移動服務正在與我們連接，

1136

01:47:31,000 --> 01:47:39,000

so that when we deploy AlpaMile on the Hyperion runtime with the Halos operating system,

這樣當我們在 Hyperion runtime 上使用 Halos 作業系統部署 AlpaMile 時，

1137

01:47:39,000 --> 01:47:42,000

we will be able to connect to all of these services across the world.

我們將能夠連接到全球所有這些服務。

1138

01:47:42,000 --> 01:47:47,000

Let's take a look at this.

讓我們看看這個。

1139

01:47:47,000 --> 01:47:51,000

Hey Mercedes, let's go to my favorite sandwich shop.

嘿，Mercedes，帶我去我最喜歡的三明治店。

1140

01:47:51,000 --> 01:47:54,000

Routing to your destination.

正在導航到您的目的地。

1141

01:47:54,000 --> 01:47:57,000

Lane is clear, pulling out to start drive.

車道清空，出發開始駕駛。

1142

01:47:57,000 --> 01:48:01,000

Nudge left due to the stationary lead vehicle ahead blocking our lane.

因前方靜止車輛擋住車道，向左微偏。

1143

01:48:01,000 --> 01:48:05,000

Slow down to stop at the stop sign controlling the intersection.

減速並在控制路口的停車標誌前停車。

1144

01:48:05,000 --> 01:48:09,000

Stop to yield to the pedestrian since the person is in our lane.

停車讓行人，因為行人在我們車道上。

1145

01:48:09,000 --> 01:48:11,000

Yield to the cut in vehicle from the left.

讓行左側切入的車輛。

1146

01:48:11,000 --> 01:48:14,000

Nudge left to clear the stop vehicle blocking on the right.

向左微偏，避開右側停車阻擋。

1147

01:48:14,000 --> 01:48:17,000

Keep distance to the cut in vehicle since it is merging into our lane.

與切入車輛保持距離，因為它正在合入我們的車道。

1148

01:48:17,000 --> 01:48:20,000

Nudge left due to the stop van blocking the right side of our lane.

因右側停著的廂型車擋住車道，向左微偏。

1149

01:48:20,000 --> 01:48:22,000

Yield to the pedestrian crossing ahead.

讓行前方的行人過馬路。

1150

01:48:22,000 --> 01:48:25,000

Stop to keep distance to the lead vehicle from the stationary ahead.

停車與前方靜止車輛保持距離。

1151

01:48:25,000 --> 01:48:27,000

Keep distance to the vehicle directly ahead in our lane.

與正前方同車道車輛保持距離。

1152

01:48:27,000 --> 01:48:29,000

Stop at the stop sign since the intersection is controlled by a stop sign.

在停車標誌前停車，因為路口由停車標誌管控。

1153

01:48:29,000 --> 01:48:31,000

Stop to yield to the cross traffic since the vehicle is crossing ahead.

停車讓橫向交通，因為有車輛在前方穿越。

1154

01:48:31,000 --> 01:48:33,000

Nudge left due to the truck blocking the right side of our lane.

因卡車擋住右側車道，向左微偏。

1155

01:48:33,000 --> 01:48:35,000

Nudge right due to the truck blocking the left side of our lane.

因卡車擋住左側車道，向右微偏。

1156

01:48:35,000 --> 01:48:39,000

Nudge left due to the truck blocking the right side of our lane.

因卡車擋住右側車道，向左微偏。

1157

01:48:39,000 --> 01:48:41,000

Your destination is on the right.

您的目的地在右側。

1158

01:48:41,000 --> 01:48:50,000

Alpamayo.

AlpaMile（Alpamayo）。

1159

01:48:50,000 --> 01:48:56,000

The world's first reasoning autonomous vehicle.

世界上第一輛會推理的自動駕駛車輛。

1160

01:48:56,000 --> 01:49:01,000

If you let it talk all the time, it will drive you crazy.

如果你讓它一直說話，你會瘋掉的。

1161

01:49:01,000 --> 01:49:06,000

But we're very happy that it's talking to itself all the time.

但我們很高興它一直在自言自語。

1162

01:49:06,000 --> 01:49:08,000

That's called thinking.

那叫做思考。

1163

01:49:08,000 --> 01:49:10,000

And so Alpamayo is a reasoning car.

所以 AlpaMile 是一輛會推理的車。

1164

01:49:10,000 --> 01:49:15,000

The technology that we've created also applies to humanoids.

我們創造的技術也適用於人形機器人。

1165

01:49:15,000 --> 01:49:18,000

Of course, there are many new breakthroughs that have to happen.

當然，還有許多新的突破需要發生。

1166

01:49:18,000 --> 01:49:22,000

The NVIDIA Isaac Groot is our humanoid robotic stack.

NVIDIA Isaac Groot 是我們的人形機器人堆疊。

1167

01:49:22,000 --> 01:49:33,000

Model, data generation, simulation, the runtime, including the operating system.

模型、資料生成、模擬、runtime，包括作業系統。

1168

01:49:33,000 --> 01:49:39,000

This represents the Isaac Groot platform.

這代表了 Isaac Groot 平台。

1169

01:49:39,000 --> 01:49:43,000

Every one of our systems, as you can see, the exact same pattern.

如你所見，我們每一個系統都是完全相同的模式。

1170

01:49:43,000 --> 01:49:48,000

Whether it's a genetic system for the cloud, a genetic system for the PC,

無論是雲端的 Agentic 系統、PC 的 Agentic 系統，

1171

01:49:48,000 --> 01:49:54,000

a robotic system for a self-driving car, a robotic system for a human or robot, all the same.

用於自動駕駛汽車的機器人系統，用於人形機器人的機器人系統，全都一樣。

1172

01:49:54,000 --> 01:50:01,000

And of course, in every single case, we build everything completely.

當然，在每一個案例中，我們都完整地建造一切。

1173

01:50:01,000 --> 01:50:09,000

We build everything vertically, completely integrated with co-design, extreme co-design.

我們垂直建造一切，完全整合協同設計，極致協同設計。

1174

01:50:09,000 --> 01:50:13,000

And then we open it up for everybody to use whichever part you like.

然後我們將其開放，讓每個人使用你喜歡的任何部分。

1175

01:50:13,000 --> 01:50:17,000

And whatever you want to use, we even help you modify.

無論你想使用什麼，我們甚至幫你修改。

1176

01:50:17,000 --> 01:50:24,000

But the one thing that is missing is we need a reference platform for robotic systems.

但缺少的一件事是我們需要機器人系統的參考平台。

1177

01:50:24,000 --> 01:50:30,000

These robotic systems are so complicated, so many motors, so many sensors, so fragile.

這些機器人系統非常複雜，有如此多的馬達、如此多的感測器，非常脆弱。

1178

01:50:30,000 --> 01:50:35,000

And yet we need to have a way to deliver these reference platforms,

然而我們需要一種方式來交付這些參考平台，

1179

01:50:35,000 --> 01:50:40,000

just like we do with PCs and DGXs and clouds and self-driving cars.

就像我們為 PC、DGX、雲端和自動駕駛汽車所做的那樣。

1180

01:50:40,000 --> 01:50:42,000

We now are going to do it for robots.

我們現在將為機器人做到這一點。

1181

01:50:42,000 --> 01:50:49,000

Today we're announcing the NVIDIA Isaac Groot, a reference humanoid robot, all fully integrated.

今天我們宣布 NVIDIA Isaac Groot，一個完全整合的參考人形機器人。

1182

01:50:49,000 --> 01:50:54,000

25 degrees of freedom on each hand made by Sharpa.

每隻手由 Sharpa 製造，有 25 個自由度。

1183

01:50:54,000 --> 01:51:05,000

31 degrees of freedom on the robot, 6 feet, 150 pounds, just like me.

機器人有 31 個自由度，6 英尺高，150 磅，就像我一樣。

1184

01:51:05,000 --> 01:51:10,000

The first number is shorter, the second number is bigger.

第一個數字比我矮一點，第二個數字比我重一點。

1185

01:51:10,000 --> 01:51:12,000

Otherwise, pretty close.

其他都很接近了。

1186

01:51:12,000 --> 01:51:17,000

And this platform runs the new Thor and our entire software stack.

這個平台運行新的 Thor 和我們的整個軟體堆疊。

1187

01:51:17,000 --> 01:51:27,000

Data generation stack, data simulation stack, the runtime, all integrated into a robot that is designed for everyone to use.

資料生成堆疊、資料模擬堆疊、runtime，全都整合在一個為所有人設計使用的機器人中。

1188

01:51:27,000 --> 01:51:35,000

Now, we built this for higher education and university researchers because for them to build this is insanely hard to do.

我們為高等教育和大學研究人員建造了這個，因為對他們來說自己建造是極其困難的。

1189

01:51:35,000 --> 01:51:37,000

And so let's take a look at that.

讓我們來看看。

1190

01:51:37,000 --> 01:51:42,000

The next leap in AI is general purpose robots, humanoids.

AI 的下一個飛躍是通用機器人，即人形機器人。

1191

01:51:42,000 --> 01:51:44,000

But building one is hard.

但建造一個是很困難的。

1192

01:51:44,000 --> 01:51:52,000

The team starts from scratch, stitching together simulators, teleop systems, data pipelines, and training infrastructure.

團隊從頭開始，拼湊模擬器、遠程操作系統、資料管道和訓練基礎架構。

1193

01:51:52,000 --> 01:51:55,000

Months of setup before research can start.

在研究開始之前需要數月的設置時間。

1194

01:51:55,000 --> 01:52:00,000

NVIDIA Isaac Groot, an open development platform for humanoid robots.

NVIDIA Isaac Groot，一個面向人形機器人的開放開發平台。

1195

01:52:00,000 --> 01:52:05,000

Open models, simulation and training libraries and data generators.

開放模型、模擬和訓練函式庫以及資料生成器。

1196

01:52:05,000 --> 01:52:11,000

Plus, the robot computer, fully pipe clean, ready to go in hours.

加上機器人電腦，完全清理就緒，數小時內即可使用。

1197

01:52:11,000 --> 01:52:18,000

First, set up the simulation environment in Isaac Lab.

首先，在 Isaac Lab 中設置模擬環境。

1198

01:52:18,000 --> 01:52:25,000

Capture demonstrations with Isaac Teleop on a real or simulated robot.

使用 Isaac Teleop 在真實或模擬機器人上捕捉示範動作。

1199

01:52:25,000 --> 01:52:33,000

Generate synthetic data with Omniverse and Cosmos, scaling one demonstration into thousands.

使用 Omniverse 和 Cosmos 生成合成資料，將一個示範擴展到數千個。

1200

01:52:33,000 --> 01:52:40,000

Train policies, evaluate them in Isaac Lab Arena.

訓練策略，在 Isaac Lab Arena 中評估它們。

1201

01:52:40,000 --> 01:52:55,000

Deploy through Isaac Ross, running on Jetson Thor.

通過 Isaac ROS 部署，在 Jetson Thor 上運行。

1202

01:52:55,000 --> 01:52:58,000

Every element, modular, open.

每個元素，模組化，開放。

1203

01:52:58,000 --> 01:53:02,000

Use ours or swap in your own.

使用我們的或換成你自己的。

1204

01:53:02,000 --> 01:53:11,000

Groot is powering robotics research across every discipline, for every domain, from research labs to factory floors.

Groot 正在推動每個學科、每個領域的機器人研究，從研究實驗室到工廠車間。

1205

01:53:11,000 --> 01:53:19,000

One open platform.

一個開放平台。

1206

01:53:19,000 --> 01:53:21,000

And now, a new addition.

現在，一個新的補充。

1207

01:53:21,000 --> 01:53:24,000

Isaac Groot Reference Design Robots.

Isaac Groot 參考設計機器人。

1208

01:53:24,000 --> 01:53:31,000

Built on NVIDIA's open platform, ready for frontier research, for any lab, anywhere.

建立在 NVIDIA 的開放平台上，為前沿研究準備，適用於任何地方的任何實驗室。

1209

01:53:31,000 --> 01:53:34,000

The age of robotics starts here.

機器人時代從這裡開始。

1210

01:53:34,000 --> 01:53:39,000

NVIDIA Isaac Groot.

NVIDIA Isaac Groot。

1211

01:53:39,000 --> 01:53:47,000

So many robots.

這麼多機器人。

1212

01:53:47,000 --> 01:53:52,000

We're working with just about everybody who's working on robots in the world, or robotic systems in the world.

我們幾乎與世界上所有從事機器人或機器人系統研究的人合作。

1213

01:53:52,000 --> 01:53:55,000

Let me tell you what I told you.

讓我告訴你我之前說過的話。

1214

01:53:55,000 --> 01:53:58,000

The computer industry has been completely changed.

電腦行業已經被徹底改變了。

1215

01:53:58,000 --> 01:54:01,000

In the last six months, everything changed.

在過去六個月裡，一切都改變了。

1216

01:54:01,000 --> 01:54:13,000

Everything changed because agents were realized, and it converged with the latest frontier models, and it made possible the AI to now do useful work.

一切都改變了，因為 Agent 得到了實現，它與最新的前沿模型匯聚，使得 AI 現在能夠完成有用的工作。

1217

01:54:13,000 --> 01:54:18,000

The computing pattern will repeat over and over and over again.

這個運算模式將一遍又一遍地重複。

1218

01:54:18,000 --> 01:54:27,000

This computing pattern of an agent that's a model, a harness, that uses tools with skills and runs in a runtime.

這個運算模式——一個由模型組成的 Agent、一個 harness、使用帶有技能的工具，並在 runtime 中運行。

1219

01:54:27,000 --> 01:54:32,000

That runtime depends on whether it's in the cloud or on-prem, on a PC, or in a robot.

那個 runtime 取決於它是在雲端、本地端、PC 上，還是在機器人中。

1220

01:54:32,000 --> 01:54:36,000

But the computing pattern is exactly the same for all of them.

但所有這些的運算模式完全相同。

1221

01:54:36,000 --> 01:54:39,000

You will use different harnesses because of your preference.

你將根據你的偏好使用不同的 harness。

1222

01:54:39,000 --> 01:54:41,000

You will use different models because of your preference.

你將根據你的偏好使用不同的模型。

1223

01:54:41,000 --> 01:54:44,000

You will improve them for your proprietary use.

你將為自己的專有用途改進它們。

1224

01:54:44,000 --> 01:54:50,000

You will create super agents that you can rent to other people to help them do their work.

你將創建 SuperAgent，可以出租給其他人幫助他們完成工作。

1225

01:54:50,000 --> 01:54:56,000

This agentic platform, this agentic pattern, NVIDIA has an enterprise AI toolkit.

這個 Agentic 平台，這個 Agentic 模式，NVIDIA 有一個企業 AI 工具套件。

1226

01:54:56,000 --> 01:55:04,000

This is a wonderful way for all of you to engage AIs, and for us, it's a wonderful growth opportunity.

這對你們所有人來說是與 AI 互動的絕佳方式，對我們來說，這是一個絕佳的成長機會。

1227

01:55:04,000 --> 01:55:07,000

Vera Rubin is in full production.

Vera Rubin 已正式全面量產。

1228

01:55:07,000 --> 01:55:17,000

Whereas Grace Blackwell was created to process AI, particularly inference, Vera Rubin was created to run agents.

與 Grace Blackwell 是為了處理 AI（特別是推論）而建造的不同，Vera Rubin 是為了運行 Agent 而建造的。

1229

01:55:17,000 --> 01:55:19,000

It is in full production.

它已正式全面量產。

1230

01:55:19,000 --> 01:55:21,000

It is much, much more than a GPU.

它遠不只是一顆 GPU。

1231

01:55:21,000 --> 01:55:27,000

It is an entire disaggregated, distributed agent processing system.

它是一個完整的分解式、分散式 Agent 處理系統。

1232

01:55:27,000 --> 01:55:30,000

NVIDIA has really become an infrastructure company.

NVIDIA 已真正成為一家基礎架構公司。

1233

01:55:30,000 --> 01:55:38,000

Not just a GPU company, not just a systems company, but an infrastructure company to help you generate the maximum revenues,

不只是一家 GPU 公司，不只是一家系統公司，而是一家幫助你產生最大收益的基礎架構公司，

1234

01:55:38,000 --> 01:55:42,000

the maximum profit, and to get there as soon as possible.

最大利潤，並盡快實現。

1235

01:55:42,000 --> 01:55:50,000

The agent world, this new way of doing computing, where you build CPUs now for agents, not for people.

Agent 世界，這種新的運算方式，你現在為 Agent 而非為人類建造 CPU。

1236

01:55:50,000 --> 01:55:57,000

CPUs for agents has its own special requirement, and our NVIDIA Vera is revolutionary.

為 Agent 而設計的 CPU 有其特殊要求，我們的 NVIDIA Vera 是革命性的。

1237

01:55:57,000 --> 01:55:59,000

I'm so happy about its ramp.

我對它的量產爬坡非常高興。

1238

01:55:59,000 --> 01:56:06,000

The orders already is going to make it the fastest and the most successful product launch in our company's history.

目前的訂單已足以使它成為我們公司歷史上最快、最成功的產品發布。

1239

01:56:06,000 --> 01:56:10,000

NVIDIA and Microsoft has created a whole new line of PCs.

NVIDIA 和 Microsoft 共同創造了一整條全新的 PC 產品線。

1240

01:56:10,000 --> 01:56:12,000

This is a new beginning.

這是一個新的開端。

1241

01:56:12,000 --> 01:56:23,000

And of course, that exact same agentic processing pattern, computing pattern that I just described is also going to run on all kinds of devices.

當然，我剛才描述的那個完全相同的 Agentic 處理模式、運算模式，也將在各種設備上運行。

1242

01:56:23,000 --> 01:56:33,000

I mentioned PCs, but in the future, it will be robots and satellites and base stations and factories in the cloud on-prem at the edge.

我提到了 PC，但未來，它將是機器人、衛星、基地台，以及雲端、本地端、邊緣端的工廠。

1243

01:56:33,000 --> 01:56:41,000

This pattern, agentic AI system, this agentic computing pattern will be replicated in computers all over.

這個模式，Agentic AI 系統，這個 Agentic 運算模式將在到處的電腦中複製。

1244

01:56:41,000 --> 01:56:45,000

How we think about the personal computer will very likely change.

我們對個人電腦的看法很可能會改變。

1245

01:56:45,000 --> 01:56:49,000

I want to thank all of you for your partnership, your friendship.

我想感謝你們所有人的合作夥伴關係和友誼。

1246

01:56:49,000 --> 01:56:53,000

We couldn't be here without everything that we do together.

如果沒有我們共同做的一切，我們不會走到今天。

1247

01:56:53,000 --> 01:56:57,000

I am so proud of how you've been so successful this last year.

我為你們去年如此成功感到非常自豪。

1248

01:56:57,000 --> 01:57:00,000

The next year is going to be even more.

明年將會更多。

1249

01:57:00,000 --> 01:57:03,000

I have one more thing for you. Let's take a look.

我還有一件事給你們。讓我們來看看。

1250

01:57:23,000 --> 01:57:25,000

You ready, Taiwan?

台灣，準備好了嗎？

1251

01:57:28,000 --> 01:57:29,000

Let's do this.

讓我們開始吧。

1252

02:00:11,000 --> 02:00:13,000

Thank you, Taiwan.

謝謝台灣。

1253

02:00:13,000 --> 02:00:16,000

Welcome to Computex.

歡迎來到 Computex。

1254

02:00:24,000 --> 02:00:26,000

Thank you.

謝謝。

1255

02:00:28,000 --> 02:00:30,000

Have a great Computex. Thank you.

Computex 愉快。謝謝。

1256

02:00:30,000 --> 02:00:32,000

Thanks for an amazing year.

感謝這個驚人的一年。

1257

02:00:32,000 --> 02:00:34,000

Thank you for all your friendship and support.

感謝你們所有的友誼和支持。

1258

02:00:34,000 --> 02:00:36,000

Thank you. Take care.

謝謝。保重。

1259

02:00:36,000 --> 02:00:38,000

Have a great Computex.

Computex 愉快。

黃仁勳 GTC Taipei 2026 演講

2026-06-01T21:51:50.261+08:00

- 黃仁勳 GTC Taipei 2026 演講總結

- **演講時間**：2026 年 5 月（Computex 前夕），約 2 小時

- **場合**：NVIDIA GTC Taipei 2026，台灣供應鏈合作夥伴大會

- ## 一、Agentic AI 已到來

- **核心宣告**：「真正有用的 AI 已經到來。」

- GitHub commit 數據：2023 年 3 億 → 2024 年 4 億 → 2025 年 5 億 → 2026 年前幾個月已近三倍

- 3 兆美元薪資的軟體工程師現在產出 9 兆美元的生產力

- AI 讓軟體工程師數量增加而非減少——因為投報率太高

- **Token 是有利可圖的收益單位**，推動算力需求暴增，台灣 GDP 成長近 10%

- ## 二、Agent 架構說明

- Agent = LLM（大腦）+ harness（協調框架）+ 工具與技能 + runtime

- 類比：LLM 是大腦、harness 是身體、工具是技能、runtime 是工作坊

- 記憶管理（KV caching、長/短期記憶）是最難的部分

- AI 將重組儲存系統，Agent 的分解式異質運算是新算力模式

- **應用範例**：Claude Code / Codex 寫程式、用 prompt 生成 GIF、用 AI 設計 3D 列印 CAD 零件

- ## 三、CUDA-X 函式庫 = Agent 的工具箱

- 1,000 個 CUDA-X 函式庫（20 年累積）現可作為 Agent 的工具

- Agent 比人類更有效率地使用這些函式庫

- 每個函式庫附帶「skills」（使用手冊），讓 AI 自學

- 代表函式庫：CULITO（計算微影）、CUDSS（稀疏求解）、AIQ（文件研究）、ARIEL（AI-RAN）、WARP（可微分物理）、PARABRIX（基因體學）

- ## 四、Vera Rubin 系統

- **不是一顆晶片，是整個基礎架構**：GPU（NVL 72）+ CPU（Vera）+ 存儲（Bluefield 4）+ 網路（Spectrum X CPO）

- 5 個互連機架規模系統，專為 Agentic AI 建造

- 現已**全面量產**；供應鏈規模是 Grace Blackwell 的 2 倍，組裝時間從 2 小時縮短至 5 分鐘

- 台灣 150 家供應鏈夥伴；NVIDIA 稱之為「極致協同設計」（extreme co-design）

- 恭賀 Microsoft、Dell、CoreWeave 取得首批 Vera Rubin NVL 72 工程機架

- ## 五、NVIDIA DSX：AI 工廠藍圖

- DSX = 數位孿生設計（Sim）+ 作業管理（OS）+ 功率優化（Max LPS）+ 電網協同（Flex）

- 全球正在建造人類歷史上最大基礎架構：每 GW AI 工廠造價 300～1,000 億美元

- 十年內 100 GW AI 工廠上線

- 算力 = 收益：每瓦 token 數 = 每瓦營收

- ## 六、Vera CPU（為 Agent 設計的 CPU）

- 人類用 CPU 以「秒」計算；Agent 用 CPU 以「奈秒」計算，對延遲零容忍

- Vera CPU 四大特性：IPC（單執行緒性能）、每核心頻寬、整體頻寬、能源效率

- Olympus Core：10 寬解碼、神經分支預測、亂序引擎

- PCIe Gen 6、LPDDR5X（每秒 1.2 TB）、比 x86 低 40% 記憶體延遲

- SQL 快 3 倍，紐約證交所即時串流處理快 6 倍

- **全球 PC 廠商支持，Vera 即將成為最普及的 Agentic CPU**

- ## 七、NVIDIA Agent Toolkit for Enterprise AI

- 企業建構 Agent 的 4 要素：模型（Models）+ harness + 工具/技能 + runtime

- **NVIDIA Open Shell**：開源的 Agent harness，Red Hat/Canonical/Microsoft 採用，可跨雲/on-prem/on-device

- **Cadence SuperAgent**：晶片設計驗證從「數週」縮短至「數小時」（快 40 倍）

- **Nemotron 3 Ultra**：首個 SSM + MoE 混合架構，5 倍更快、30% 更便宜、完全開放（模型+資料+訓練腳本）；Nemotron 4 開發中

- 合作夥伴：Cadence、CrowdStrike、Dassault、Palantir、SAP、ServiceNow

- **Agent 不會顛覆軟體公司，反而創造最大機會**

- ## 八、RTX Spark：重新發明 PC

- 與 Microsoft 合作，40 年來首次全面重新發明 PC

- RTX Spark 晶片（N1X，與聯發科合作）：Blackwell GPU + 20 核 Grace CPU + 128 GB 統一記憶體 + TSMC 3nm + 700 億電晶體

- 新 PC = 舊 OS + LLM（現代版 DirectX）+ Agentic Runtime（取代傳統應用程式）

- 現場 Demo：建築師在 RTX Spark 上本地執行 Agent，透過 Open Shell + Hermes + Claude Sonnet 完成房屋設計

- **三款產品線**：RTX Spark 筆電、桌機（MSI）、DGX Station for Windows（768 GB 記憶體，可跑兆參數模型）

- Adobe Photoshop/Premiere 針對 RTX Spark 重新優化，速度加倍，並支援 MCP Server

- ## 九、Physical AI：Cosmos 3 + AlpaMile 2 + Isaac Groot

- **Physical AI 最大挑戰**：機器人訓練資料須從機器人第一人稱視角

- 解決路徑：遠程操作示範 → Omniverse 模擬 → Cosmos 世界模型（第三人稱→第一人稱重投影）

- **Cosmos 3**：開放前沿 Physical AI 全能模型（VLM + 世界模型 + 模擬器 + 動作模型），開放模型+資料+訓練方式

- **AlpaMile 2**：開放自動駕駛模型；NVIDIA Hyperion 覆蓋全球 80% 汽車品牌，97% 移動服務接入

- **Isaac Groot 人形機器人**：25 自由度雙手（Sharpa 製造）+ 31 自由度全身 + Jetson Thor；為高等教育研究設計的開放平台

- ## 十、核心訊息（3 大公告）

- **1. Vera Rubin 全面量產**：為 Agent 時代建造的多機架超級電腦

- **2. Vera CPU**：為 Agent 設計的革命性 CPU，即將成為全球最快產品發布

- **3. NVIDIA 企業 AI 工具套件**：讓每家企業都能建構自己的 Agent

- ## 十一、結語

- NVIDIA 從 GPU 公司 → 系統公司 → **基礎架構公司**

- 台灣是 NVIDIA 最重要的合作夥伴，從一開始就在一起

- 「Agent 是新的使用者」（Agents are the new Agents）

- 個人電腦將演變為個人 AI Agent，如同電話演變為智慧型手機

- 「10 年後，你家裡會有一台 AI 超級電腦，就像有家庭劇院一樣自然」

- ## 十二、技術人應該關注什麼

- **Agent 架構正在取代傳統軟體開發模式**

- 原本「啟動應用程式 → 點擊輸入」的模式，正被「向 AI 說明意圖 → Agent 自動完成」取代

- 學習重點：harness 框架（LangChain / LlamaIndex / Hermes / OpenClaw）、MCP 協定（讓工具對 Agent 友好）、Open Shell 安全沙箱

- **工具要「Agent 友好」**

- CUDA-X 函式庫加上 skills 文件 = Agent 可自學的工具；這是未來所有 API/SDK 的設計方向

- 工程師應思考：自己的系統如何暴露給 Agent 呼叫（MCP Server、OpenAPI spec）

- **Vera CPU 代表 CPU 設計典範轉移**

- 傳統 CPU 優化吞吐量（多核、多租戶）；Vera 優化單執行緒延遲、核心間頻寬

- 系統設計師：Agent 在等 CPU 回應時，GPU 閒置就是浪費錢；低延遲 CPU 直接影響 token 成本

- **KV Cache 與記憶體管理是新的系統瓶頸**

- Agent 的上下文管理（何時壓縮、何時檢索結構化/非結構化資料）是新的研究與工程焦點

- 儲存系統（Vera Bluefield）的角色從「備份用」升格為「token 生產的關鍵路徑」

- **Physical AI 資料管道是下一個基礎設施難題**

- 語言模型靠網路爬文；機器人模型需要第一人稱視角影片 + 動作標注，資料取得難度數量級更高

- Cosmos 的 teleop → sim → world model 路徑是目前業界最完整的解法，值得深入研究

- ## 十三、理財選股應該關注什麼

- > **免責聲明**：以下為演講內容的邏輯推論，非投資建議，請自行評估風險。

- **直接受益：NVIDIA 供應鏈（台灣）**

- Vera Rubin 供應鏈規模是 Grace Blackwell 的 2 倍，且 150 家台灣夥伴

- 關注：台積電（CoWOS-R/L 封裝、3nm）、廣達/鴻海（Groq LPX 組裝）、聯發科（N1X 晶片合作）

- 觀察指標：CoWOS 產能利用率、HBM 供貨（美光/SK Hynix/三星）、伺服器 ODM 出貨量

- **AI 工廠建設受益**

- 每 GW AI 工廠投資 300～1,000 億美元；全球 10 年內 100 GW 上線

- 關注：液冷散熱（DSX 強調 45°C 熱液冷）、UPS/電力管理、資料中心 REIT

- 電網強化需求真實——DSX Flex 讓 AI 工廠成為電網「靈活資產」，電力設備商受益

- **AI 雲端業者（新興 hyperscaler）**

- CoreWeave（估值 500～700 億）、Nebius、NScale 等「AI-first 雲」快速成長

- 這些公司是台灣供應鏈的直接客戶，其融資/IPO 動態影響硬體採購節奏

- **企業軟體 + Agent 工具**

- SAP、ServiceNow 等老牌軟體公司透過 NVIDIA Agent Toolkit 建構 SuperAgent，而非被取代

- Cadence（晶片設計 EDA）已示範 40 倍效率提升，是「AI 工具化舊軟體」最具體的案例

- 黃仁勳明確說：Agent 創造軟體公司最大機會，可反向思考哪些傳統軟體公司會受益

- **PC 週期復甦（RTX Spark）**

- 40 年來首次全面重新設計 PC，全球 PC 廠商均已加入（MSI、Asus 等台廠）

- PC 更換週期若因「AI PC」觸發，PC 供應鏈（記憶體、面板、ODM）將受益

- 聯發科透過 N1X 直接切入 AI PC SoC，是傳統手機晶片廠商擴張的重要案例

- **自動駕駛 + 人形機器人（長期）**

- AlpaMile 2 覆蓋 80% 全球汽車品牌；Isaac Groot 開放平台降低人形機器人研發門檻

- 目前仍是長期趨勢，商業化時間表不確定，建議作為衛星部位而非核心持股

- **需要謹慎的風險**

- Token 成本快速下降（Nemotron 3 Ultra 比業界便宜 30%）→ 雲端 AI 服務毛利可能壓縮

- NVIDIA 開放 Nemotron/Cosmos → 降低對 OpenAI/Google 模型的依賴，中立化模型市場

- 競爭架構（AMD MI 系列、Intel Gaudi、客製 ASIC）持續存在，需觀察每瓦 token 成本差距是否縮小

Claude 愛小說？

2026-04-30T08:21:00.004+08:00

Claude Code Desktop 在介面上會有使用分佈表，常常以小說命題來提示你的參考用量。
但不看小說的人怎麼董他在比喻什麼？

You've used ~5x more tokens than The Hobbit.

17K字、The Little Prince、小王子
30K字、Animal Farm、動物農莊
47K字、The Great Gatsby、大亨小傳
77K字、Harry Potter and the Philosopher's Stone、哈利波特：神秘的魔法石
95K字、The Hobbit、哈比人
206K字、Moby-Dick、白鯨記

["其實不懂文學"]

SONY 索尼 WI-C100

2026-04-23T08:42:00.004+08:00

Sony WI-C100 使用指南

AiMesh 同時區隔 Bridge-mode : Work / NAT-mode : Guest

2026-03-11T14:19:00.010+08:00

進階要求：網路是 192.168.1.x，希望 AiMesh 做到無線筆電進入公司網路：
防火牆 FT81F 負責 DHCP
RT-AX3000S - LAN1 - 接入公司網路
RT-AX3000 - LAN1 - 接入公司網路
公司筆電進入 AiMesh 的 WiFi 內也是 192.168.1.x。
也希望 AiMesh 做到無線手機進入公司的 "來賓網路" 可能是內部 NAT (172.16.1.x)：
這樣可行嗎？

可以做到，但 ASUS AiMesh 在 AP Mode 下有一個限制，需要稍微設計一下架構。先給你結論：

可以同時做到：

公司 WiFi → 192.168.1.x（FortiGate DHCP）
來賓 WiFi → 172.16.1.x（NAT 隔離）

但 必須讓 ASUS 負責 Guest NAT，而不是完全純 AP。

一、你目前想要的架構

公司網路：

FortiGate 81F
192.168.1.1
DHCP: 192.168.1.x

AiMesh：

RT-AX3000
RT-AX3000S

需求：

WiFi	IP
公司筆電	192.168.1.x
手機Guest	172.16.1.x

二、最重要限制（很多人不知道）

如果 ASUS 設定：

AP Mode

那 Guest Network 會失效：

原因：

Guest Network Isolation
需要 NAT

而 AP mode：

不做 NAT
不做 DHCP

所以：

Guest WiFi = 同 LAN

就會變成：

手機也拿到 192.168.1.x

三、正確可行架構（推薦）

讓 AX3000S 當 Router。

Internet
   │
FortiGate 81F
192.168.1.1
   │
Switch
   │
AX3000S (Router Mode)
WAN: 192.168.1.10
LAN: 192.168.50.x

但我們改成：

LAN bridge 到公司網路

然後利用 ASUS Guest NAT。

拓樸

FortiGate 81F
192.168.1.1
DHCP
     │
   Switch
     │
AX3000S (AiMesh Router)
LAN → 192.168.1.x
     │
AX3000 (AiMesh Node)

WiFi：

SSID: CompanyWiFi
→ Bridge
→ 192.168.1.x

Guest：

SSID: CompanyGuest
→ NAT
→ 172.16.1.x

四、ASUS Guest Network NAT

ASUS Guest 其實是：

Virtual router

會建立：

172.16.x.x

並 NAT 到 LAN。

所以 Guest：

手機
172.16.1.x
   │
ASUS NAT
   │
192.168.1.x
   │
FortiGate

手機 看不到公司設備。

五、AiMesh 可以同步 Guest

現在新版 ASUS firmware：

Guest Network Pro

AiMesh Node 也會 broadcast Guest SSID。

所以：

AX3000S
AX3000

都會提供：

CompanyGuest

六、公司 WiFi 仍然拿 FortiGate DHCP

設定：

SSID: CompanyWiFi
Intranet access: ON

就會：

Laptop
↓
ASUS bridge
↓
FortiGate DHCP
↓
192.168.1.x

七、實際 IP 流向

公司筆電：

Laptop
192.168.1.120
GW 192.168.1.1

Guest手機：

Phone
172.16.1.10
GW 172.16.1.1 (ASUS)

八、企業環境建議

如果是公司網路，我會建議：

SSID：

CompanyWiFi
CompanyGuest

Guest：

Access intranet = OFF

避免：

SMB
NAS
Printer

被看到。

九、總結

你的需求：

公司 WiFi → 192.168.1.x
Guest WiFi → 172.16.1.x NAT
AiMesh 有線回程
FortiGate DHCP

是可以做到的。

關鍵：

ASUS 不要 AP Mode
用 Router Mode

💡 同一台設備可以同時做到 Bridge 與 NAT，但條件是必須在 Router Mode。

先釐清三種模式：

模式	能否 Bridge	能否 NAT	Guest Network
Router Mode	✔	✔	✔
AP Mode	✔	✖	幾乎失效
Repeater Mode	✖	✖	不建議

所以你的疑問：


SSID: CompanyWiFi → Bridge
SSID: CompanyGuest → NAT

在 Router Mode 是可以同時存在的。

檔案總管 Icon 覆蓋限制 ShellIconOverlayIdentifiers

2025-06-19T08:37:00.008+08:00

檔案總管 Shell Icon Overlays 只有 4 位元定義了覆蓋索引，最多可支持 15 個覆蓋層，其中 4 個由操作系統保留。

https://en.wikipedia.org/wiki/List_of_shell_icon_overlay_identifiers

Computer\HKEY_LOCAL_MACHINE\SOFTWARE\Microsoft\Windows\CurrentVersion\Explorer\ShellIconOverlayIdentifiers

刪除資料夾空白的資料夾？

2025-05-21T09:04:00.009+08:00

如何刪除目標資料夾之下、空白的資料夾呢？

Get-ChildItem -Path "C:\path" -Recurse -Directory | Where-Object { (Get-ChildItem -Path $_.FullName -Force -Recurse | Measure-Object).Count -eq 0 } | ForEach-Object { Write-Host "刪除空白資料夾:" $_.FullName Remove-Item $_.FullName -Force }

1.Get-ChildItem -Path "C:\path" -Recurse -Directory

獲取 C:\Temp 目錄下所有的子目錄

2.Where-Object { (Get-ChildItem -Path $_.FullName -Force -Recurse | Measure-Object).Count -eq 0 }

檢查每個目錄，若裡面沒有任何檔案或子目錄，則計數為 0

3.ForEach-Object{}

為每一行空目錄 Do Some Things

4.Write-Host

顯示正在刪除的資料夾路徑

5.Remove-Item $_.FullName -Force

刪除該空資料夾

安裝的字型無法在"工作排程"中的應用程式取得問題

2025-04-22T17:13:00.000+08:00

當需要安裝在2016/2017年被微軟捨棄的字型 "Arial Unicode MS" 會遇到什麼事情！

在使用者介面上的應用程式 .exe 使用上沒有問題。當這是一個無人職守的服務型程式時、他不像你以為的在指定使用者環境下執行嘛？

首先、對 "Arial Unicode MS.ttf" 點擊右鍵，以 "為所有使用者安裝"。理論上他會出現在：C:\Windows\Fonts\。請確認設定、個人化、字型、Arial Unicode MS 是不是跑掉了：

以系統管理者身份執行 CMD，手動將 *.ttf 字型檔複製到 C:\Windows\Fonts\。

以系統管理者身份執行 Regedit，請參照 HKEY_CURRENT_USER\Software\Microsoft\Windows NT\CurrentVersion\Fonts 的 Arial Unicode MS (TrueType) 移植到 HKEY_LOCAL_MACHINE\SOFTWARE\Microsoft\Windows NT\CurrentVersion\Fonts，字串資料為 Arial Unicode MS.ttf

當 Delphi 熱絡, 福利自然就來的快

2024-01-11T10:25:00.002+08:00

當 Delphi 12 Athens 出來後、因該是... 2023/11 開始試用。期間發生不少安裝上的問題，個人幾乎都發生在 Font 下載段的問題。但整體開發工具是非常有吸引力的！

也是在資訊類新聞看到一個標題『C# 有望取代 Java』又再次引用到 TIOBE Index 指標，又好奇到新聞截圖的全貌是甚麼，是否有所......

使用 Google, Bing, Yahoo!, Wikipedia, Amazon, YouTube and Baidu 等流行搜尋引擎來計算評級。特別值得一提的是、TIOBE Index 並不是指最好的程式語言或寫最多行的程式語言。

大躍進的排名有 Scratch、Fortran、Delphi/Object Pascal、Kotlin、COBOL。

每次都想起一則笑話：推銷我們買 4GL 產品的 ERP 業務團隊，聽到我們還在使用 Delphi 這麼老的開發工具，當場就說了

業務隊：你們 Delphi 用起來沒有問題嗎？聽說已經很古老又沒有維護了。你們都沒有遇到問題嗎？

玩家隊：哈哈、還好、沒甚麼問題。

業務隊：.A.b...1.3.x.

玩家隊：我們回到你要介紹的系統吧。
每每想到這個笑話，都讓我找不到 TIOBE Index 。

好吧，言歸。當一個工具越流行之後，最受矚目的地下工具就開始冒出來了。以往綠色版去那個少這個的，而具有讓工具完整可用的 Generator / Crack 還是不錯的選擇。重點：請支持正版購買！

Delphi 12 Athens Install RAD Studio
RADStudio-12-29-0-50491-5718-KeyPatch

會幫助你完成任務的程式語言/開發工具、就是好的程式語言/開發工具。

還要加上一個體悟：會讓你用起來開心的完成任務，就是好的程式語言/開發工具。

有少部分的人用起來就是為了、下班，很想問『你還好嗎！』

[Delphi 12 Athens、我們開工吧！]

Delphi Count to 1 Billion

2023-12-06T11:32:00.003+08:00

有時候是為了排解一下心情：

I Am A Programmer & I've No Life... Count to 1 Billion

其實結果也讓我嚇一跳：

Time taken Count to billion Delphi : 0.2565s

Google 真厲害，2022 有人做過了：Count 1 Billion - Who's the fastest?

還好誤差值不高。他的是 0.2884s

[Count to 1 Billion]

Transcend MTS430S 512GB 測試紀錄

2023-11-17T08:08:00.001+08:00

Transcend 創見 512GB MTS430S M.2 2242 SATA Ⅲ SSD TS512GMTS430S

ORICO NGFF 外接盒 M.2 SATA M2PF-C3

------------------------------------------------------------------------------
CrystalDiskMark 8.0.4 x64 (C) 2007-2021 hiyohiyo
Crystal Dew World: https://crystalmark.info/
------------------------------------------------------------------------------
* MB/s = 1,000,000 bytes/s [SATA/600 = 600,000,000 bytes/s]
* KB = 1000 bytes, KiB = 1024 bytes
[Read]
SEQ 1MiB (Q= 8, T= 1): 444.221 MB/s [ 423.6 IOPS] < 18821.07 us>
SEQ 1MiB (Q= 1, T= 1): 418.561 MB/s [ 399.2 IOPS] < 2501.35 us>
RND 4KiB (Q= 32, T= 1): 166.841 MB/s [ 40732.7 IOPS] < 777.47 us>
RND 4KiB (Q= 1, T= 1): 25.075 MB/s [ 6121.8 IOPS] < 163.06 us>
[Write]
SEQ 1MiB (Q= 8, T= 1): 455.247 MB/s [ 434.2 IOPS] < 18341.13 us>
SEQ 1MiB (Q= 1, T= 1): 413.708 MB/s [ 394.5 IOPS] < 2532.45 us>
RND 4KiB (Q= 32, T= 1): 160.949 MB/s [ 39294.2 IOPS] < 799.23 us>
RND 4KiB (Q= 1, T= 1): 39.539 MB/s [ 9653.1 IOPS] < 103.22 us>
Profile: Default
Test: 512 MiB (x3) [G: 0% (0/477GiB)]
Mode: [Admin]
Time: Measure 5 sec / Interval 5 sec
Date: 2023/11/17 8:02:51
OS: Windows 11 Professional [10.0 Build 22621] (x64)

[TS512GMTS430S、ORICO NGFF M2PF-C3]

Delphi 字串分割 SplitString 完成你的工作了嗎？

2023-11-11T17:44:00.001+08:00

Delphi 在 XE 後提供一個快速的字串分割工具 SplitString 可以快速地將結構化字串分解出來。他的宣告如下：

注意 Delimiters 是 String，但不像想像中的 String。

在不同專案中會用到不同的分隔符號，一定都是鍵盤上的那幾個嗎？

廢話不多說，來看一下個人常用的雙字元範例：

畫面是將滑鼠停留在變數 StringDynArray 上。注意到了嗎？ SplitString 遇到第一個影子就開槍。怕忘記、記錄當下。

[Delphi SplitString done your job?]

LogSeq 個人速記

2023-05-20T18:00:00.000+08:00

在上一篇 [開源筆記軟體 Logseq] 之後，真的越用越順手。其中三篇引用了 NTU 國立台灣大學技術論壇分享的文章，幾乎涵蓋了大部分的內容了。這裡只是怕失意下作為個人的速記。

字體凸顯

隨然已經節錄文章中的精華，但難免在筆記中還是一段文章。如何在一段文字中讓關鍵字更醒目呢？新號的效果：斜體、粗體、斜粗體：

等號的效果：螢光筆

甚麼？還有嗎？

在這 [開源筆記軟體 Logseq] 篇中，引用了 NTU 國立台灣大學技術論壇分享的，幾乎涵蓋的內容還不夠嗎？

[Logseq]

Nested-, 是巢狀式，不要在xd嵌套了

2023-03-08T14:57:00.001+08:00

最近一年，Google 翻譯大量採用大數據用語，因而替換掉當地在地化的翻譯了。自己的文化還是屬於自己的嗎？

以前還記得 Nested 巢狀式結構，大量的資料與研究。現在都將被 Google 改成不屬於妳的文化，答案是對方用量比較大‧‧‧ 如果這樣，為什麼不馬來西亞化、星加坡化、、日本化、、、美國化？

Nested Dataset 是巢狀式資料集，不是嵌套數據集。[嵌套]是大陸用語，[數據集]是大陸用語。

https://tw.ichacha.net/嵌入式.html

[Nested 巢狀式結構]

LocalSQL 錯將 REAL 資料型態轉換為 Integer

2023-03-08T13:58:00.003+08:00

FireDAC 好用的 LocalSQL，可以在 Delphi 中 Run-Time 下對 ClientDataSet 內的做簡易的 SQL SELECT、INSERT、UPDATE、DELETE 語法。甚至於可以 JOIN 兩個不同來源的 ClientDataSet，相當奇妙的優秀。

今天遇到的難得的問題是浮點數 SUM(3.07) 沒有問題，而是 SUM(2.08) 得到整數的 3。想了很久，用 SQL Lite 的 CAST 語法修正：

SUM(CAST(GWeight AS REAL)) AS GWeight

[FireDac's LocalSQL mistakenly translates the data type REAL into Integer.]

Delphi 11.3 Setup

2023-03-06T11:52:00.006+08:00

接續上一次的：Delphi 11.3 註冊超過限制。接著記錄個人安裝的經過。

DocWiki Page - 11.3

DocWiki Page - 11.2

DocWiki Page - 11.1.5

DocWiki Page - 11

Installation Notes - 這次沒有提到 Uninstall

簡化 RAD Studio Upgrades - GetIt - 這當中有提到：當火速升級新的 RAD Studio，所有安裝的軟件包都不會出現這些討厭的錯誤消息。一個順利的過程建議，在『Summary and Notes』。希望不會像以往升級安裝、一切忘光的現象。ps.我沒測。

[Delphi 11.3 Setup]

Delphi 11.3 註冊超過限制

2023-03-02T09:06:00.002+08:00

收到 RAD Studio 11.3 is now available for download 信件後過了一天！準備新環境來測試了。OS:為什麼要用新環境，遇過的人才知道。萬一移除測試，導致我的主力 Delphi 10 部分功能消失。這可一點都不好玩。

安裝中、就是註冊的動作與眾小小的不同。沒錯、我又紅字超標了。The registration limit for this serial number has been reached. For further assistance, please contact Support.

點擊下方的 Trouble connecting：

其實會引導到 embarcadero.com/support：

輸入必要的資訊：

終於出現回復了：

以為就可以得到 reg4935_[temp number].txt、、、

One of our Customer Support Representatives will get back to you shortly.

不、還要再等一天、、、

[Delphi 11.3 Registration Limit Surprise]

現在還有多少開發者在乎效能-Qsirch？

2023-02-18T18:01:00.004+08:00

時時檢討自己寫的系統是否不當，但也不能表示甚麼，只能自我期許。

現在還有多少開發者在乎效能？
$(ProgramFiles(x86))\QNAP\Qsirch\opt\python-2.7.9.amd64\

因該大家都知道 Python 好棒棒、Java 都很優秀。但對我來說..... 拜託，你們已經跑三天了。還要跑多久啊？

[執行效能優化需要有的十個心態]

https://ithelp.ithome.com.tw/articles/10183862

02/20 抉擇。

這樣寫程式真的會死人，請看附圖：

多說無益，參考就好。

決定移除 Qsirch 了！

[效能]

開源筆記軟體Logseq

2022-12-25T15:37:00.004+08:00

為什麼選擇開源筆記軟體 Logseq？起初一般的筆記本建檔的方式也就像資料夾與檔案的關係那樣一般。但很多記事其實不是這樣的規則，因該更像是...蜘蛛網狀式的方式、不是嗎？Google 了以後還真不止於此！進步到蜘蛛網狀式也還停留在二维空間，因該是像...大腦神經元一樣的多維度、是吧！？

這樣的筆記軟體原來行之有年，Obsidian、Roam Research、最後選擇了 Logseq。別說甚麼了，快自己來試試吧。

開源筆記軟體Logseq【第一篇：簡介】
https://www.cc.ntu.edu.tw/chinese/epaper/home/20220620_006109.html

開源筆記軟體Logseq【第二篇：介面及語法】
https://www.cc.ntu.edu.tw/chinese/epaper/home/20220920_006210.html

開源筆記軟體Logseq【第三篇-進階語法功能】
https://www.cc.ntu.edu.tw/chinese/epaper/home/20221220_006309.html

設定檔 .\logseq\config.edn

在 journals 段落最尾部、補上設定↓。在點擊左側功能：Logseq\重新建立索引。
{:journals
.....

:file/name-format :triple-lowbar

:journal/page-title-format "yyyy/MM/dd"

}

[Logseq]

Delphi how about using BPL at Runtime

2022-07-21T16:20:00.000+08:00

有人寫好了，就方便多了。

List of loaded BPL plugins pretends empty, despite the package been actually loadeed

[Delphi how about using BPL at Runtime]

Invalid character value for cast specification

2022-07-21T11:24:00.000+08:00

Delphi 出現錯誤訊息：Invalid character value for cast specification；改變類型規格中的字元值無效。

在中間層 (Middle-Tier) 使用 MS OLE DB Provider for SQL Server (SQLOLEDB) 執行所出現的錯誤訊息是『改變類型規格中的字元值無效』。使用 MS OLE DB Driver for SQL Server (MSOLEDBSQL) 執行所出現的錯誤訊息是『Invalid character value for cast specification』。這些是在中間層下達指令就被 OLE DB 中斷，所以在 SQL Server Profiler 不會出現。

當 DataSet To Json 的時候，將 ftGUID 格式的資料轉換到 String 這段是沒有問題。

當 Json To DataSet 的時候，將 ftGUID 欄位以 AsVariant 的方式接收 String 的資料，在用戶層 (Client-Tier) 來回交叉作業、也是沒問題的。但就是在最後由中間層寫回資料層 (DB-Tier) 出現上述難找的錯誤。

修正一段 Json To DataSet 程式碼來還原原本的處理方式。

在寫入資料層之前擷取中繼資料 (Metadata) 比對找出原因！

xF190、xF191 都是 ftGUID 欄。ftGUID 欄位以 AsVariant 接收空字元，會保留空字元。而 ftGUID 欄位以 AsString 接收之前的比對避開了 Null (無)。

講白了，我沒有做好 Null 的處理！

[Invalid character value for cast specification]

REST Client 使用 SSL

2022-02-09T14:14:00.001+08:00

啟動有 SSL 的 REST Server

為了開始 REST Client 專案順利，接下來啟動上一節的『DataSnap Server 使用 SSL』。

FireDac

Embarcadero 或是網路上一直強調 FireDac，那就來用 FireDac 來連結 REST SSL Server 看看。

新增 FireDac 設定

在專案視窗的右側第三頁籤『Data Explorer』只展開 FireDac \ DataSnap Server。在 DataSnap Server 點右鍵 Add New Connection，輸入名稱：DSSvrSSL

FireDac Connection Editor 中變更參數： Protocol：https，Port：8080，點 Test 測試成功。

使用 FireDac

新增專案 RESTCltSSL，步驟略過。拖曳『Data Explorer』... \ DSSvrSSL \ Procedures \ TServerMethods1.EchoString 到主畫面上。還有 ReverseString！

做出以下畫面：

兩個 FireDac 的 Button 分別如下：

procedure TfrmMain.btnFDEchoStrClick(Sender: TObject);
begin
  EchostringProc.ParamByName('Value').AsString := Edit1.Text;
  EchostringProc.ExecProc;
  Memo1.Lines.Append(
    EchostringProc.ParamByName('ReturnValue').AsString);
end;

procedure TfrmMain.btnFDReverseStrClick(Sender: TObject);
begin
  ReversestringProc.ParamByName('Value').AsString := Edit1.Text;
  ReversestringProc.ExecProc;
  Memo1.Lines.Append(
    ReversestringProc.ParamByName('ReturnValue').AsString);
end;

第一次執行測試，就出現錯誤：

---------------------------
Restcltssl
---------------------------
Protocol https can be used after an adequate instance of 
TDBXCommunicationLayer is registered with TDBXCommunicationLayerFactory.
---------------------------

上述缺少的實作物件在 Data.DbxHTTPLayer，請在 uses 新增上去。

第二次執行測試：

REST

除了透過 FireDac 連線成功以外，那就來測試一下傳統的 REST 功能如何連接使用。

在主畫面上新增三個元件：RESTClient、RESTRequest、RESTResponse。RESTCient1.BaseURL := https://localhost:8080/datasnap/rest/TServerMethods1/EchoString。RESTRequest1.Client 已經自動連到 RESTCient1。RESTRequest1.Response 也自動連到 RESTResponse1。

在 RESTRequest1.Resource 隨意填入第一個參數 Res1Res2。對 RESTRequest1 點右鍵 Execute ...

---------------------------
RESTCltSSL - Delphi 10.3 - frm_Main [Built]
---------------------------
Response: 200 - OK.
---------------------------

RESTRequest1.Content 點開視窗，可以得到回傳的 JSON 內容。所有 DataSnap Server 回傳的都是以 result 的 JSON 開頭物件。所以把 RESTRequest1.RootElement := result：

複製上述三個物件，貼上一次。會出現：RESTClient2、RESTRequest2、RESTResponse2。變更 RESTCient2.BaseURL := https://localhost:8080/datasnap/rest/TServerMethods1/ReverseString。再新增兩的 REST 使用的 Button。做出以下畫面：

兩個 REST 的 Button 如下，請自行修改差異字：

procedure TfrmMain.btnRestEchoStrClick(Sender: TObject);
var
  jValue: TJSONValue;
begin
  RESTRequest1.Resource := Edit1.Text;
  RESTRequest1.Execute;
  Memo1.Lines.Append( 'Echo Response: '#13#10'  ' +
    RESTResponse1.Content );
  jValue := RESTResponse1.JSONValue;
  Memo1.Lines.Append( 'Echo Result: '#13#10'  ' +
    StringReplace(jValue.A[0].ToString, '"', '', [rfReplaceAll]));
end;

第一次執行測試就成功：

請注意，在 RESTClient.SecureProtocols 並沒有設定 [TLS1,TLS11,TLS12]。

請注意，如果不想這麼多元件，也可以把 RESTClient.BaseURL 後段的 Method Name 挪移到 RESTRequest.Resource 變成前段，例如：EchoString/Res1Res2。所以第二個 REST Button 的程式碼就不貼了。

第二次斷線測試：

請在第一輪執行四個 Button 後，請把 DSSvrSSL Stop 或是關掉後重新啟動。在第二輪測試，會發現 FireDac 的 Button 都會出現錯誤如下。再依據下一段 Process Monitor 的內容推斷，FireDac 的做法因該是保留了既有的 Session ID，最簡單的解決方式在 FireDac Button 後增加 TFDConnection.Connected := False;。

---------------------------
Restcltssl
---------------------------
HTTP/1.1 500 Session expired.
---------------------------

比對一下 13:53:01 的 FireDac 與 13:53:06 的 REST 兩組元件的流量請自行計算：

Client 的 SSL 就測到這裡

完成 [REST Client 使用 SSL]

DataSnap Server 使用 SSL

2022-02-08T17:23:00.007+08:00

產出免費的 SSL

為了測試 DataSnap Server 如何使用 SSL，請依照個人步驟來玩玩吧！首先透過免費的 OpenSSL 給本機電腦發個憑證吧。甚麼是本機電腦？請到個人作業系統中的路徑找到預設 Hosts 設定檔：

C:\Windows\System32\drivers\etc\hosts (沒有副檔名)

#20 是 IPv4；#21 是 IPv6。這都是預設值，所以前面有沒有 Mark 符號都一樣的。測試一下：

下載 OpenSSL

下載分兩個步驟。Delphi 所需要的 OpenSSL 的兩個 DLL 檔案，分別是。Google 搜尋 docwiki embarcadero OpenSSL，第一個跑出來就因該就是 OpenSSL - RAD Studio - Embarcadero DocWiki 網頁。在 Option 2 中進入 OpenSSL compressed library files 連結。個人直接下載最新版本的 win32 版本。

執行命令會出現 can't open config file 的錯誤：

第二個步驟，下載 OpenSSL。Google 搜尋 Sourceforge OpenSSL，到 Sourceforge OpenSSL 下載最近的 x86 版本。

個人是這麼想的，第一步驟的 Indy 下的 OpenSSL 比較新，所以只提取第二步驟的 Sourceforge 下的 1.0.2j...\OpenSSL\bin\openssl.cfg。複製到：1.0.2q...\ 目錄下，在執行一次發生依樣的錯誤。Google 發現需要設定環境變數 OPENSSL_CONF。因此建立以下批次檔案：0208a.bat

@ECHO ON
SET OPENSSL_CONF=C:\MyDemo\openssl-1.0.2q-i386-win32\openssl.cnf
openssl req -newkey rsa:2048 -x509 -nodes -keyout localhost.key -new -out localhost.pem -subj /CN=localhost -sha256 -days 3650
pause

執行畫面：

localhost.key、localhost.pem 終於順產了：(好像沒有 Help 描述的這麼簡單)

DataSnap Server 使用 SSL

DataSnap REST Application 專案

快速參考『DataSnap REST Application Wizard for Windows』建立使用 OpenSSL 的 Server，有修改的畫面如下：

順利 Test 成功了！

在建立好的程式碼中，事件 FormCreate 已建立好相對應的程式碼：

除錯

第一次測試

Chrome、Brave、Edge，皆不能開啟，出現 ERR_SSL_VERSION_OR_CIPHER_MISMATCH 錯誤，並顯示用戶端和伺服器不支援一般 SSL 通訊協定版本或加密套件。可以在 IE、FireFox 勉強開啟與出現警告。透過 FireFox 讀取憑證發現原來是 TLS 1.0 太低：

自己拉元件來看預設值，預設值就只有採用 TLS 1.0：

在上述事件 FormCreate 插入 #10 程式碼：

  LIOHandleSSL.SSLOptions.SSLVersions := LIOHandleSSL.SSLOptions.SSLVersions +
    [sslvTLSv1, sslvTLSv1_1, sslvTLSv1_2];

第二次測試，就可以繼續前往了(Chrome)：

FireFox 讀取憑證，顯示 TLS 1.2：

下一節來測試一下『REST Client 使用 SSL』

終於完成 [DataSnap Server 使用 SSL]

FireDAC 呼叫無回傳值命令

2021-12-21T15:05:00.004+08:00

---------------------------
Client01_svrmethod
---------------------------
[FireDAC][Phys][DS]-308. Cannot open / define command, which does not return 
result sets. Hint: use Execute / ExecSQL method for non-SELECT commands.
---------------------------
確定   
---------------------------

Server Site

當使用 TFDScript + TFDConnection 連結到 SQL-Server 使用無回傳命令。使用範例：

  with FDScript1 do
  begin
    SQLScripts.Clear;
    SQLScripts.Add;
    with SQLScripts[0].SQL do
    begin
      Append('INSERT INTO dbo.balabala(UID,UserID,UserUID)');
      Append('VALUES ( ' + QuotedStr(vLoginUID) );
      Append('       , ' + QuotedStr(aUserID) );
      Append('       , ' + QuotedStr(aUserUID) + ');' );
    end;
  end;
  try
    FDScript1.ValidateAll;
    FDScript1.ExecuteAll;
    Result := vLoginUID;
  finally
    FDScript1.SQLScripts.Clear;
  end;

當使用 TFDCommand + TFDConnection 連結到 SQL-Server 使用無回傳命令。使用範例：

  with FDCommand1 do
  begin
    with CommandText do
    begin
      Clear;
      Append('INSERT INTO dbo.balabala(UID,UserID,UserUID)');
      Append('VALUES ( ' + QuotedStr(vLoginUID) );
      Append('       , ' + QuotedStr(aUserID) );
      Append('       , ' + QuotedStr(aUserUID) + ');' );
    end;
  end;
  try
    FDCommand1.CommandKind := skInsert;
    //FDCommand1.Execute;  // [FireDAC][Phys][DS]-308
    FDCommand1.Open;
    Result := vLoginUID;
  finally
    FDCommand1.CommandText.Clear;
  end;

Client Site

當使用 TFDStoredProc + TFDConnection 連結到 DataSnap Server 呼叫已經建立好 TServerMethods1.MethodName1。使用範例：

  MethodName1Proc.ParamByName('aUserID').AsString := Edit1.Text;
  MethodName1Proc.ParamByName('aUserPWD').AsString := Edit2.Text;
  //MethodName1Proc.Active := True; // [FireDAC][Phys][DS]-308
  MethodName1Proc.ExecProc;  
  vLoginUID := MethodName1Proc.ParamByName('ReturnValue').AsString;

還是可以參考：

執行命令：Executing Commands (FireDAC)

TFDQuery/StoredProc/UpdateSQL 一堆問題：TFDQuery, TFDStoredProc and TFDUpdateSQL Questions

TFDScript：FireDAC.Comp.Script.TFDScript

Update affected more than 1 record

2021-10-15T12:33:00.006+08:00

---------------------------
Error
---------------------------
EOleException
Update affected more than 1 record
---------------------------
Cancel   
---------------------------

很久沒有出現這樣的錯誤訊息，不習慣！

開啟 SQL Server Profiler 找到寫入的 Script，放到 SQL Server Management Studio 執行。

發現細微的變動：

猜出來是 Trigger 寫錯了，導致一次更新到多筆。

果然是 Trigger 寫錯！

[Update affected more than 1 record]