The AI Infrastructure Revolution: Why NVIDIA Matters / AI基础设施革命：为何英伟达举足轻重

The AI Infrastructure Revolution: Why NVIDIA Matters

AI基础设施革命：为何英伟达举足轻重

In the summer of 2023, something remarkable happened at Jensen Huang's keynote presentation at the SIGGRAPH conference. The NVIDIA CEO, known for his black leather jacket and energetic stage presence, paused mid-demo to address the elephant in the room. The company's market capitalization had just crossed the $1 trillion mark—a figure that would have seemed fantastical just three years earlier. "We're at the iPhone moment of AI," Huang declared, his words carrying the weight of both a prediction and a confession. The confession was implicit: even NVIDIA, the company that had spent decades building the chips that would enable this moment, had not fully anticipated how quickly or how completely the artificial intelligence revolution would transform the technology landscape—and with it, the fate of a company that started by making video game graphics cards.

2023年夏季的SIGGRAPH大会上，发生了一件值得关注的事情。英伟达CEO黄仁勋（Jensen Huang）以他标志性的黑色皮夹克和充满活力的舞台风格，在演示中途停下来回应了房间里那头"房间里的大象"。此时，公司的市值刚刚突破1万亿美元大关——这个数字在三年看来还显得荒诞不经。"我们正身处AI的iPhone时刻，"黄仁勋宣布，他的话语承载着预测与坦白双重分量。隐含的坦白是：即便是英伟达这家花费数十年打造这些芯片的公司，也没有完全预料到人工智能革命会以如此迅速且彻底的方式改变技术格局——以及由此改变一家从视频游戏显卡起家的公司的命运。

This book is the story of that transformation—of how a company known primarily to gamers and computer enthusiasts became the backbone of the most significant technological shift since the invention of the internet. It's a story about hardware and software, about vision and luck, about ecosystems and moats. Most importantly, it's a story that reveals why understanding NVIDIA matters not just for tech enthusiasts or investors, but for anyone who wants to understand the infrastructure that will increasingly shape our economic, social, and even political reality.

本书讲述的正是这场转型——一家主要为游戏玩家和电脑爱好者熟知的公司，如何成为自互联网诞生以来最重要的技术变革的支柱。这是一个关于硬件与软件、关于远见与运气、关于生态与护城河的故事。更重要的是，这个故事揭示了理解英伟达为何重要——不仅对科技爱好者或投资者重要，对任何想要理解这种将日益塑造我们经济、社会乃至政治现实的底层设施的人来说，都至关重要。

What Made This Moment Possible

是什么让这一刻成为可能

To understand why NVIDIA became so central to the AI revolution, we need to understand something fundamental about what artificial intelligence actually requires from computing hardware. Modern AI systems, particularly the large language models and neural networks that have captured the world's attention, are fundamentally about pattern recognition at massive scale. They learn by processing billions of examples, adjusting countless internal parameters until they can recognize patterns in new data they've never seen before.

要理解英伟达为何在AI革命中变得如此核心，我们需要理解人工智能实际上对计算硬件有什么根本需求。现代AI系统，特别是引起全球关注的大型语言模型和神经网络，本质上是在进行大规模的模式识别。它们通过处理数十亿个样本进行学习，不断调整无数内部参数，直到能够识别从未见过的新数据中的模式。

This process of learning—called training in machine learning terminology—is computationally intensive in a very specific way. Unlike traditional computing tasks that often require a processor to perform complex sequences of operations one after another (called serial processing), training neural networks requires performing the same relatively simple mathematical operations simultaneously on millions or billions of pieces of data. This is called parallel processing, and it's where GPUs excel.

这个被称为训练（training）的学习过程，以一种非常特殊的方式消耗大量计算资源。与传统计算任务通常需要处理器一个接一个地执行复杂序列操作（称为串行处理）不同，训练神经网络需要在数百万甚至数十亿个数据片段上同时执行相同的相对简单的数学运算。这被称为并行处理（parallel processing），而这正是GPU的强项所在。

The story of why GPUs are so well-suited for AI begins, paradoxically, with their original purpose: rendering graphics for video games. Video games present a unique computational challenge. A game world must respond in real-time to player input, rendering millions of pixels dozens of times per second to create the illusion of continuous motion. Each pixel's color depends on complex calculations involving light sources, textures, object positions, and viewing angles. And here's the crucial insight: while each pixel's calculation is independent of the others, all millions of calculations must happen simultaneously to produce a coherent image.

GPU之所以如此适合AI，这个故事说起来颇具讽刺意味，因为它始于GPU的原始用途：渲染视频游戏图形。电子游戏提出了一种独特的计算挑战。一个游戏世界必须实时响应玩家输入，每秒数十次渲染数百万像素，以创造连续运动的错觉。每个像素的颜色取决于涉及光源、纹理、物体位置和视角的复杂计算。关键洞见在于：虽然每个像素的计算相互独立，但为了产生连贯的图像，所有数百万次计算必须同时进行。

This requirement gave rise to a fundamentally different kind of processor architecture. A traditional CPU, like the Intel processor in a laptop, is designed as a general-purpose workhorse. It can handle a wide variety of tasks, switching between them rapidly to give the appearance of simultaneity. Modern CPUs typically have between 4 and 16 "cores"—individual processing units that can work on different tasks. A GPU, by contrast, contains thousands of smaller, simpler cores. Each core is less powerful than a CPU core individually, but together they can process vast amounts of data in parallel.

这一需求催生了一种根本不同的处理器架构。传统的CPU，如笔记本电脑中的英特尔处理器，被设计为通用工作主力。它可以处理各种任务，在不同任务之间快速切换，给人生成同时进行的错觉。现代CPU通常有4到16个"核心"——可以处理不同任务的独立处理单元。相比之下，GPU包含数千个更小、更简单的核心。每个核心单独来看不如CPU核心强大，但它们一起可以并行处理海量数据。

This architectural difference is visualized in the comparison below:

这种架构差异可以从下图中得到直观理解：

```mermaid graph TB subgraph CPU["Traditional CPU Architecture"] C1[Core 1] --> CS[Control Unit] C2[Core 2] --> CS C3[Core 3] --> CS C4[Core 4] --> CS CS --> ALU[Arithmetic Logic Unit] end

subgraph GPU["GPU Architecture"] subgraph SM1["Streaming Multiprocessor"] G1[Core 1] --> SH1[Shared] G2[Core 2] --> SH1 G3[Core 3] --> SH1 G4[Core 4] --> SH1 end subgraph SM2["Streaming Multiprocessor"] G5[Core 5] --> SH2[Shared] G6[Core 6] --> SH2 G7[Core 7] --> SH2 G8[Core 8] --> SH2 end subgraph SMn["..."] GN1[Core N-3] --> SHn[Shared] GN2[Core N-2] --> SHn GN3[Core N-1] --> SHn GN4[Core N] --> SHn end end

style CPU fill:#e1f5fe style GPU fill:#fff3e0 ```

Figure 1.1: CPU vs. GPU Architecture Comparison 图1.1：CPU与GPU架构对比

The implications of this architecture for AI were not immediately obvious. When NVIDIA's founders Jen-Hsun Huang, Chris Malachowsky, and Curtis Priem started the company in 1993, they had one goal: create graphics chips that could deliver realistic video game experiences. The parallel processing architecture that would later prove so crucial for AI was originally designed to solve a different problem entirely—how to render triangles, apply textures, and calculate lighting fast enough to keep up with the demands of increasingly sophisticated games.

这种架构对AI的影响最初并不明显。当英伟达的创始人黄仁勋、克里斯·马拉科夫斯基（Chris Malachowsky）和柯蒂斯·普里姆（Curtis Priem）于1993年创立公司时，他们只有一个目标：创造能够提供逼真游戏体验的图形芯片。后来被证明对AI至关重要的并行处理架构，最初是为了解决一个完全不同的问题而设计的——如何足够快速地渲染三角形、应用纹理和计算光照，以跟上日益复杂的游戏需求。

The Serendipitous Discovery: CUDA and the Birth of GPU Computing

偶然的发现：CUDA与GPU计算的诞生

The bridge between gaming graphics and artificial intelligence was built gradually, beginning in the mid-2000s. In 2006, NVIDIA introduced CUDA (Compute Unified Device Architecture), a parallel computing platform and programming model that allowed developers to use NVIDIA GPUs for general-purpose computing tasks—beyond just graphics. CUDA was initially met with skepticism. Why would anyone want to program a graphics chip to do non-graphics work? The GPU was seen as a specialized tool, good for games and maybe some scientific simulations, but certainly not for the kind of general-purpose computing that CPUs were designed for.

游戏图形与人工智能之间的桥梁是逐渐搭建起来的，起始于2000年代中期。2006年，英伟达推出了CUDA（Compute Unified Device Architecture，统一计算设备架构），这是一个并行计算平台和编程模型，允许开发者使用英伟达GPU进行通用计算任务——不仅仅是图形处理。CUDA最初遭到了质疑。为什么有人会想用图形芯片来做非图形工作？GPU被视为一种专业工具，适合游戏，也许还有某些科学模拟，但绝不是为CPU设计的那种通用计算。

NVIDIA, however, saw something others didn't. The company invested heavily in CUDA, making it easier to use, spreading it to universities for research, and creating software libraries that made GPU computing accessible to non-specialists. This