图形革命：GPU的诞生与英伟达的创立

第一章：图形革命——GPU的诞生与英伟达的创立

Chapter 1: The Graphics Revolution — The Birth of GPU and the Founding of NVIDIA

当1993年黄仁勋（Jensen Huang）站在加州弗里蒙特的一间小办公室里时，他或许已经预见到这一天将会改变整个计算产业的轨迹。但即便他本人，恐怕也难以想象这家初创公司将在三十年后成为人工智能时代最重要的基础设施供应商之一。GPU——这个最初仅为绘制视频游戏中的三维图形而设计的专用处理器，如今已经成为驱动深度学习、自动驾驶、科学研究乃至元宇宙构建的核心引擎。理解GPU如何诞生、它与CPU的本质区别是什么、以及英伟达如何从一家濒临破产的图形芯片公司成长为市值超过万亿美元的科技巨头——这些问题的答案，将帮助我们理解当代人工智能革命的底层逻辑。

When Jensen Huang stood in a small office in Fremont, California in 1993, he may have foreseen that this day would change the trajectory of the entire computing industry. But even he himself probably couldn't have imagined that this startup would become one of the most important infrastructure suppliers of the artificial intelligence era three decades later. GPU—the dedicated processor originally designed solely for rendering three-dimensional graphics in video games—has now become the core engine driving deep learning, autonomous driving, scientific research, and even the construction of the metaverse. Understanding how the GPU was born, what its fundamental differences are from CPUs, and how NVIDIA grew from a near-bankrupt graphics chip company into a tech giant with a market cap exceeding one trillion dollars—answers to these questions will help us understand the underlying logic of the contemporary AI revolution.

要理解GPU的出现，我们首先要回到计算机图形学的蛮荒时代——那个在今天看来几乎无法想象的、连基本的几何线条都难以在屏幕上流畅绘制的年代。

To understand the emergence of GPU, we must first return to the wilderness era of computer graphics—an age that seems almost unimaginable today, when even basic geometric lines couldn't be rendered smoothly on screen.

从向量到像素：计算机图形学的艰难起步

From Vectors to Pixels: The Difficult Beginnings of Computer Graphics

计算机图形学的发展史，本质上是一部与硬件限制不断斗争的历史。1963年，伊凡·苏泽兰（Ivan Sutherland）在麻省理工学院Lincoln实验室开发的Sketchpad系统，被广泛认为是现代计算机图形学和用户界面设计的开山之作。Sketchpad能够用光笔在屏幕上绘制和操作几何图形，它展示了人机交互的全新可能——但实现这一可能性的代价是：整个系统需要占据整整一个房间的设备，耗电相当于一栋小型办公楼。

The history of computer graphics is fundamentally a history of continuous struggle against hardware limitations. In 1963, Ivan Sutherland's Sketchpad system, developed at MIT's Lincoln Laboratory, is widely recognized as the foundation of modern computer graphics and user interface design. Sketchpad could use a light pen to draw and manipulate geometric shapes on screen, demonstrating entirely new possibilities for human-computer interaction—but the price of realizing this possibility was that the entire system required equipment occupying an entire room and power consumption equivalent to a small office building.

在那个时代，显示器的成像原理与今天截然不同。早期的矢量显示器（Vector Display）直接在屏幕上绘制线条——电子束按照指令在阴极射线管（CRT）上描绘路径，只有点亮的区域才会发光。这使得显示的图像具有一种独特的“透明”美感：所有线条都像是悬浮在黑暗中发光的几何框架。但这种技术的致命缺陷在于，它无法高效地填充封闭区域。当苏泽兰想在Sketchpad中显示一个实心圆时，他必须用数千条短直线去逼近圆的轮廓——每一条线都需要电子束精确“飞”过，才能在视觉上形成连续的效果。

In that era, how displays formed images was entirely different from today. Early vector displays drew lines directly on screen—the electron beam traced paths according to instructions on the cathode ray tube (CRT), with only illuminated areas glowing. This gave displayed images a unique "transparent" aesthetic: all lines appeared as luminous geometric frames suspended in darkness. But the fatal flaw of this technology was that it couldn't efficiently fill enclosed areas. When Sutherland wanted to display a solid circle in Sketchpad, he had to approximate the circle's outline with thousands of short straight lines—each line required the electron beam to precisely "fly" across, only then could it form a continuous visual effect.

随着个人电脑的兴起，另一种显示技术逐渐成为主流：光栅图形学（Raster Graphics）。与矢量显示器不同，光栅显示器将屏幕划分为一个由像素（Pixel）组成的规则网格——这就是我们今天所熟知的分辨率概念。一块分辨率为640×480的显示器，包含307,200个独立的像素点，每个像素都可以独立控制其颜色和亮度。这种架构的优势在于它可以显示任意复杂的图像，包括照片、纹理和实心填充区域；代价则是每一帧画面都需要精确控制每一个像素的状态。

As personal computers rose to prominence, another display technology gradually became mainstream: raster graphics. Unlike vector displays, raster displays divide the screen into a regular grid of pixels—the resolution concept we're familiar with today. A display with 640×480 resolution contains 307,200 independent pixels, each controllable for color and brightness. The advantage of this architecture is that it can display arbitrarily complex images, including photos, textures, and solid filled areas; the cost is that every frame requires precise control over every pixel's state.

这里出现了一个核心矛盾：图形渲染的计算需求与可用硬件资源之间的巨大鸿沟。假设我们要在640×480的分辨率下显示一个包含多边形场景的三维游戏，每秒30帧的流畅播放意味着处理器需要在每一秒钟内完成近一亿次像素级别的操作——这还不包括几何变换、光照计算、纹理映射等更复杂的处理任务。在20世纪80年代的条件下，这个计算量对于任何通用处理器来说都是不可承受之重。

Here emerges a core contradiction: the vast gap between the computational demands of graphics rendering and available hardware resources. Suppose we want to display a three-dimensional game with polygon scenes at 640×480 resolution, with smooth playback at 30 frames per second—this means the processor must complete nearly 100 million pixel-level operations every second—and this doesn't even include more complex processing tasks like geometric transformations, lighting calculations, and texture mapping. Under 1980s conditions, this computational load was unbearable for any general-purpose processor.

```mermaid graph LR A[3D Scene] --> B[Vertex Processing<br/>顶点处理] B --> C[Primitive Assembly<br/>图元组装] C --> D[Rasterization<br/>光栅化] D --> E[Fragment Processing<br/>片元处理] E --> F[Framebuffer<br/>帧缓冲] F --> G[Display Output<br/>显示输出]

style A fill:#e1f5fe style B fill:#b3e5fc style C fill:#81d4fa style D fill:#4fc3f7 style E fill:#29b6f6 style F fill:#039be5 style G fill:#01579b ```

上图展示了一个典型的三维图形渲染管线（Rendering Pipeline）的各个阶段。从场景描述到最终显示输出，每一帧画面都需要经过一系列连续的计算步骤。这些步骤中，有些天然具有并行性——比如光栅化阶段需要独立计算每一个像素的颜色——而有些则具有严格的顺序依赖——比如某些几何变换必须在前一步完成后才能进行。理解这一管线结构，是理解GPU与CPU架构差异的关键前提。

The diagram above illustrates the various stages of a typical three-dimensional graphics rendering pipeline. From scene description to final display output, every frame must go through a series of continuous computational steps. Among these steps, some have inherent parallelism—for example, the rasterization stage needs to independently calculate the color of each pixel—while others have strict sequential dependencies—for example, certain geometric transformations can only proceed after the previous step completes. Understanding this pipeline structure is the key prerequisite for understanding the architectural differences between GPU and CPU.

三维图形革命：从学术幻想到消费级应用

The 3D Graphics Revolution: From Academic Fantasy to Consumer Application

20世纪80年代末至90年代初，计算机图形学迎来了一场静悄悄的革命。这场革命的核心，是三维图形技术从实验室走向消费市场的艰难转型。

From the late 1980s to early 1990s, computer graphics underwent a quiet revolution. At its core was the difficult transition of three-dimensional graphics technology from laboratory to consumer market.

1981年，硅谷图形公司（Silicon Graphics Inc., SGI）推出了划时代的IRIS 1000图形工作站。SGI的创新在于，它专门为图形渲染设计了一套完整的硬件和软件生态系统。IRIS图形系统包含了专用几何处理芯片，可以加速矩阵运算和坐标变换；它运行的是一种叫做IRIX的UNIX变体操作系统，并配备了名为GL的革命性图形库。GL库将复杂的图形操作封装成简洁的函数调用，开发者无需了解底层硬件细节，就可以调用诸如`glBegin`、`glVertex`、`glLight`这样的接口来构建三维场景。这套系统的出现，使得高质量三维图形首次在专业领域得到广泛应用——从CAD设计到电影特效，从科学可视化到飞行模拟器。

In 1981, Silicon Graphics Inc. (SGI) launched the groundbreaking IRIS 1000 graphics workstation. SGI's innovation was designing a complete hardware and software ecosystem specifically for graphics rendering. The IRIS graphics system included dedicated geometry processing chips that could accelerate matrix operations and coordinate transformations; it ran a variant of UNIX called IRIX and featured a revolutionary graphics library called GL. The GL library encapsulated complex graphics operations into simple function calls—developers could call interfaces like `glBegin`, `glVertex`, and `glLight` to construct three-dimensional scenes without needing to understand low-level hardware details. The emergence of this system enabled high-quality three-dimensional graphics to be widely applied in professional fields for the first time—from CAD design to film visual effects, from scientific visualization to flight