英伟达第一部：GPU公司（1993-2006）

谁掌握了真相？是你吗？是你吗？是你吗？现在谁掌握了真相？

Who got the truth? Is it you? Is it you? Is it you? Who got the truth now?

是你吗？是你吗？是你吗？让我坐下。直说吧。

Is it you? Is it you? Is it you? Sit me down. Say it straight.

又一个故事即将展开。谁掌握了真相？

Another story on the way. Who got the truth?

欢迎收听《Acquired》第十季第五期，这是一档关于伟大科技公司及其背后故事与策略的播客。我是本·吉尔伯特，西雅图Pioneer Square Labs的联合创始人兼董事总经理，同时也是我们风投基金PSL Ventures的管理者。

Welcome to season 10 episode five of Acquired, the podcast about great technology companies and the stories and playbooks behind them. I'm Ben Gilbert, and I'm the cofounder and managing director of Seattle based Pioneer Square Labs and our venture fund, PSL Ventures.

我是大卫·罗森塔尔，一位常驻旧金山的天使投资人。

And I'm David Rosenthal, and I am an angel investor based in San Francisco.

我们是本期主持人。按市值计算，它是全球第八大公司。哇。当英伟达于1993年创立时，它在一个竞争激烈且利润微薄的市场上生产计算机图形芯片。当时有90家无差异化的竞争者基本都在同时做着同样的事情，而如今，他们在独立GPU市场占据了83%的份额——对于刚接触这个领域的听众来说，GPU就是图形处理器——这些产品供应给台式机和笔记本电脑。

And we are your hosts. It is the eighth largest company in the world by market cap. Dang. When NVIDIA began in 1993, it made computer graphics chips in a brutally competitive and low margin market. There were 90 undifferentiated competitors all doing basically the same thing at the same time, and yet today, they have an 83% market share of standalone GPUs, that's graphics processing units for those of you starting with us from square one, that are supplied for desktop and laptop computers.

本，你这是在讲述整个故事啊。

Ben, you're telling, like, the whole story here.

非常抱歉。抱歉。我简单提几点。不仅于此，英伟达的追随者们都知道，他们最近开创了一个全新市场——为机器学习、神经网络、深度学习提供硬件和软件开发工具，所有这些都在云端和数据中心运行，显然这正在定义整个计算领域的十年发展。

So sorry. Sorry. I'll just I'll tease a few things here. So not only that, but, of course, the followers of NVIDIA know that they recently pioneered a completely new market. The hardware and software development tools to power machine learning, neural networks, deep learning, all of this in the cloud and the data center, which obviously is proving to define this whole decade of computing.

当大卫和我开始研究时，我们意识到这完全可以写成一本惊险小说般的书。因为联合创始人兼CEO黄仁勋曾三次押上整个公司命运，每次都几乎破产。但显然，站在今天回望，那些危机都化险为夷了。

And as David and I began our research, we realized this really could be a book and like a thriller of a book since the co founder and CEO Jensen Huang really has bet the company, like the whole company, three separate times, nearly going bankrupt each time. But obviously, as we reflect back here today, that certainly did not happen.

好了，在我们要花六小时讨论他之前，先快速了解黄仁勋的精华版：这家伙以前开丰田Supra跑车，就像《速度与激情》里那种死亡机器，还差点因此丧命。

Alright. So here's everything you need to know about Jensen. The CliffsNotes before we talk for, like, six hours about him. The dude used to drive a Toyota Supra, like a fast and the furious style Yes. Like like a death machine, and he almost died.

他遭遇过一场严重车祸。

He got in, like, a huge accident.

这又让他和埃隆·马斯克多了一个相似点。

Just one more way he is like Elon Musk.

天啊，太疯狂了。

Oh, man. Crazy.

由于本期内容实在太多，我们将把机器学习的故事留到下次。今天要讲述的是英伟达从创立到崛起，推动计算机图形学和游戏革命的狂野历程。这堪称真正发明创新的典范，提醒我们工程突破确实推动世界前进。补充些背景，这个故事时间跨度大约从1993年到2000年代中后期。

Well, because we have way too much here for one episode, we'll save the stories on machine learning for next time. Today, are gonna tell the wild story of NVIDIA's founding to its rise in prominence powering the computer graphics and gaming revolution. This really is a story of, like, true invention and innovation. It reminds you that engineering breakthroughs really do push our world forward. And in saying that, just kinda set some context, this is a story that takes place from about 1993 to kind of the mid to late two thousands.

尽管英伟达在过去五年里备受瞩目，股价飙升，公司周围充满热情，但我认为黄仁勋仍然是一位被低估的首席执行官。甚至可以说，

And as hyped as NVIDIA has been, you know, over the last five years, obviously, with the stock run up and everyone's excitement around the company, I think Jensen is still an underrated CEO. Even rated

100%同意。

100%.

英伟达的多头们对他的评价。我认为黄仁勋是这样一个人：如果你了解他，你就明白我们在说什么，并对他怀有无比的敬意。但我觉得真正了解他的人还不够多。

Where the NVIDIA bulls have put him. I think Jensen is one of those people where, like, if you know about him, you know what we're talking about and you have unbelievable reverence. But I think not enough people really know.

在进入正题前，再分享一句黄仁勋的名言。这句最棒：'我的求生意志几乎超过了所有人想杀死我的意志。'太精彩了。

Just one more Jensen quote before we get into the episode. This is the best. My will to survive exceeds almost everybody else's will to kill me. Amazing.

好了，听众朋友们。现在正是感谢我们最喜爱的公司之一Anthropic的最佳时机，它已成为我们Acquired工作流程的核心部分，以及他们最新的突破性模型Claude Sonnet 4.5。

Alright, listeners. Now is a great time to thank one of our favorite companies that has become a core part of our workflow for Acquired, Anthropic, and their latest breakthrough model, Claude Sonnet 4.5.

没错。在研究这些标志性公司时，我们不断提出诸如'他们处理这种情况的方式有何独特之处？'或'那个策略有多新颖？之前有其他公司尝试过吗？'这类问题。而能够对这些问题给出深思熟虑的答案，正是当今企业在构建AI时所需要的。

Yes. As we research these iconic companies, we're constantly asking questions like, what was unique about the way they approach this situation? Or how novel was that strategy? And had any other companies tried it before? These kind of questions and the ability to produce thoughtful answers to them are exactly what today's enterprises need when building with AI.

而Claude确实能够推理并回答这些问题。

And Claude can actually reason through and answer them.

Claude Sonnet 4.5不仅是又一个模型，它是全球最优秀的编程模型，也是构建复杂智能体的最佳选择。Shopify和Netflix的工程师称其为强大的思考伙伴，并告诉我们它正在改变他们的开发速度。Canva在其部分产品中使用Claude，称其为重大飞跃。企业们对Sonnet 4.5赞不绝口。

Claude Sonnet 4.5 isn't just another model. It's the best coding model in the world and the most capable for building complex agents. Engineers at Shopify and Netflix call it their powerful thinking partner and tell us that it is transforming their development velocity. And Canva, which uses Claude for some of its products, calls it a big leap forward. Companies are loving Sonnet 4.5.

有一点已经很清楚：让一个模型擅长编程，也让它天生擅长任何分析性任务。因此，使Claude擅长重构代码库的特性，同样让它擅长梳理数千份监管文件或进行复杂的财务分析。通过Anthropic的API，Claude与企业现有工作流程无缝集成，现在具备新的记忆和上下文管理功能，使智能体能长时间运行而不丢失关键信息。

And one thing that's become clear is that making a model great at coding also makes it great at any analytical task right out of the box. So the same thing that makes Claude great at refactoring code bases also makes it great at, say, combing through thousands of regulatory documents or doing complex financial analysis. Claude integrates seamlessly with enterprises existing workflows through Anthropics API and now has new memory and context management features that let agents run longer without losing critical information.

无论你是在扩展工程团队，还是构建下一代智能应用，Claude都能与你共同思考复杂性，而不仅仅为你思考。它真正是你智慧的思考伙伴。

So whether you're scaling an engineering team or building the next generation of intelligent applications, Claude thinks through complexity with you, not just for you. It is truly your intelligent thought partner.

请访问claude.ai/acquired免费试用Claude，并享受Claude Pro三个月五折优惠。如果你想了解他们的企业服务，只需告诉他们是本和大卫推荐你的。

So head on over to claude.ai/acquired to try Claude for free and get 50 off Claude Pro for three months. And if you wanna get in touch about their enterprise offerings, just tell them that Ben and David sent you.

听众朋友们，听完本期节目后，如果你心想'天啊，我真想和人聊聊这个'，我们有好消息告诉你：你可以在acquired.fm/slack与Acquired社区其他11000名聪明成员交流。还有个新消息——如果我们上次提到这个可能是几年前了——Spotify手机应用刚刚新增了评分功能。

Listeners, after you finish this episode and you're thinking to yourself, gosh, I wish I could talk about this with people, we have good news for you. You could do that with 11,000 other smart members of the Acquired community at acquired.fm/slack. Here's a new thing. If you haven't rated or reviewed this podcast yet I think the last time we we mentioned this was like years ago. Spotify in their mobile app just added the ability to rate.

所以如果你用Spotify收听，请务必给我们留个小评分。如果在Apple Podcasts，请给我们写个评论。我们真的非常非常感激你作为听众与他人分享你的体验。好了听众们，这不是财务建议。

So if you listen in Spotify, you should totally leave us a little, rating in there. If you're on Apple Podcasts, leave us a review. We really, really, really appreciate it when you help share your experience as a listener with others. Alright, listeners. This is not financial advice.

我们可能持有节目中讨论内容的仓位。本节目仅供娱乐和信息参考。大卫，带我们进入下一环节。

We may hold positions in things we discuss on this show. This is for entertainment and informational purposes only. And David, take us in.

那么我们从1963年2月开始说起。1963年的硅谷发生了什么？让我想想。仙童半导体已经创立了，我认为。硅谷那时正处于起步阶段，但还处于早期。

So we start in February 1963. What's going on in Silicon Valley in 1963? Let's see. Fairchild had already started, I think. Silicon Valley was, like, underway, but it was early days.

但我们的故事并非始于硅谷，而是台湾？没错。在台湾岛的南部，黄健生（后美国化名为黄仁勋）出生了。他的父亲是空调公司开利的工程师。

But we start not in Silicon Valley, but in Taiwan? Yes. The southern part of the island of Taiwan with the birth of Jian Soon Huang, later Americanized to Jensen. Jensen Huang. So his dad was an engineer for the air conditioning company carrier.

哦，是啊。

Oh, yeah.

是的。就是那些建筑物上大型工业空调机组。当黄仁勋四岁时，他父亲去美国纽约参加公司培训。他被震撼了，心想：'太棒了，我希望孩子们能在这里长大，获得所有可能的机会。'

Yeah. You see those like big like industrial air conditioning units on buildings and stuff. And when Jensen is four, his dad goes on a company training to America, to New York City. And he was like, wow, you know, this is amazing. I want my kids to grow up here and to have all the opportunities that are available.

于是他回国后，四岁的黄仁勋和他大几岁的哥哥开始学英语——家里没人会说英语。母亲买了本英语词典，每天挑十个单词，严格训练两个孩子，用词典教他们英语。现在你听黄仁勋说话，他的口音从何而来？因为这与你想象的不同。

So he comes home, Jensen's four, Jensen has an older brother who's a couple years older. You know, like nobody speaks English. So his mom gets an English dictionary and picks 10 words every day, grills the two kids, like quizzes them, and teaches them English out of the dictionary. Now, you listen to Jensen, where does that accent come from? Cause it's not what you would think.

几年后全家搬到了泰国。当黄仁勋九岁时，他们终于决定送孩子去美国。父母当时还没钱移民，但找到了一所价格低廉的美国寄宿学校——肯塔基州东部乡村的奥奈达浸会学院。黄仁勋后来说，他和哥哥是该校首批外国学生，也极可能是奥奈达镇历史上第一批华人。哇。

The family ends up moving to Thailand a few years later and then when they're living in Thailand and Jensen is nine, they finally decide that this is the right time to send the kids to America. Now the parents can't move to America yet, they they don't have enough money, but they found a boarding school in America that is cheap enough that they can afford. It is called Oneida Baptist Institute and it is in Eastern Kentucky, the sticks of Kentucky. Jensen would later say that he and his brother were the first foreigners to attend this school and they're pretty sure they were the first Chinese people ever in the town of Oneida. Woah.

后来发现这所奥奈达浸会学院如此便宜的原因：它根本不是预科学校，而是问题少年改造学校。所以当九岁的黄仁勋入学时，他的室友是个刚出狱的17岁少年，身上还带着刀战中留下的七处刀伤。

Well, it turns out that the reason that this school, OBI, Oneida Baptist Institute was so cheap was it's actually not a prep school, it's a reform school. So this is a school for, like, troubled kids. It's a reform school. So Jensen's roommate, when he shows up as a nine year old, is a 17 year old kid who had just gotten out of prison and was recovering from seven stab wounds that he got in a knife fight.

经典的美式旅程就在这里。

Classic American journey right here.

令人惊讶的是，这很符合詹森的作风。尽管这个孩子比他大八岁，基本上年龄是他的两倍，背景也截然不同，但他们却成了好朋友。詹森帮他学数学，而他带詹森接触举重。所以你现在看詹森，就会觉得这家伙肌肉发达。

And amazingly, this is so Jensen. Like, they become great friends even though this kid is eight years older than him, like twice his age basically from a way different background. Jensen helps him with math and he gets Jensen into weight lifting. So you see Jensen today and you're like, that dude is jacked.

他确实肌肉发达。

He is jacked.

他从九岁就开始举重了。谈到在安妮塔的时光时，他说："现在我不太会害怕了。我不担心去没去过的地方。我能忍受很多不适。"天啊，这话说得太准了，哇。

He's been weight lifting since he was nine years old. He says about his time in Anita, you know, now I don't get scared very often. I don't worry about going places I haven't gone before. I can tolerate a lot of discomfort. Boy, does that play out Wow.

正如我们将看到的，这在他的人生中得到了印证。现在的情况其实相当棒。他和妻子洛丽已经向学校捐赠了几百万美元，那里现在成了一所了不起的机构。可以看到詹森在2020年发表了毕业演讲。我们会在资料来源里链接这个。

In his life as we will see. So it's pretty awesome actually now. He and his wife Lori have given a few million dollars to the school and it's like a amazing institution now. You can see Jensen gave the commencement address in 2020. We're gonna link to this in the sources.

这真的很棒。在OBI待了几年后，他的父母终于攒够钱能自己来美国了。他们先搬到了华盛顿州的塔科马，伟大的华盛顿州。然后他们又往南搬了一点，来到俄勒冈州波特兰的郊区。詹森和他弟弟回家和他们一起住，在那里上公立学校。

It's pretty awesome. So after a couple years at OBI, his parents are finally able to save up enough money to afford to come to The US themselves. So they move first to Tacoma, Washington, great state of Washington. And then they move a little farther south down to the suburbs of Portland, Oregon. Jensen and his brother go home, they live with them, they go to public school there.

詹森继续着他的美式成长。他迷上了乒乓球，在全国青少年锦标赛中获得第三名，还登上了《体育画报》的版面。

You know Jensen continues his his American upbringing. He gets really into table tennis. He places third in the junior nationals in, table tennis, and he gets his picture in Sports Illustrated.

不可能。

No way.

相当惊人。但他的父母继续保持着那种学术纪律，显然詹森非常聪明。他最终跳了两级然后上大学。他上了州内大学，就读于俄勒冈州立大学，就在不远处。

Pretty amazing. But his parents continue their sort of like academic discipline, and Jensen's super smart, obviously. He ends up skipping two grades and then going to college. He goes to in state college. He goes to Oregon State University, just down the road a little bit.

他到那里时才16岁左右，对吧？

And he got there when he was like 16. Right?

16岁到那里是因为他跳了两级。他热爱数学，所以决定在俄勒冈州立大学主修电气工程。结果他完全坠入了爱河——而且是多方面的。首先他爱上了电气工程，觉得这是世界上最酷的东西，成为了学校的顶尖学生之一。

Got there when he was 16 because he had skipped a couple grades. And he loves math, so he decides he's gonna major in electrical engineering at OSU. And he totally falls in love in more ways than one. The first way that he falls in love is, he just thinks, like, electrical engineering is the coolest thing in the world. Becomes one of the top students in the school.

他谈到自己会因教授们在讨论精确数字时不够严谨而生气。

He talks about how, like, he gets mad at the professors because they don't use, like, enough precision when talking about, like, exact numbers.

后来他表示尊重相反立场。一些英伟达员工称之为CEO数学——当他四舍五入所有数字时。他反思说，确实理解了教授们想表达的：只有先把握大局，细节才重要。

Which he later comes to say that he respects the opposite position. I think some of the NVIDIA employees call it CEO math when he sort of rounds all the numbers and he's like, I I reflecting back, I do understand what the professors were trying to show is like, the details only matter if you understand the big picture first.

这太詹森了。他理解员工们对他四舍五入用CEO数学的恼怒。他也欣赏精确，但大局才是关键。他第二次坠入爱河是和他的电气工程基础课实验搭档。

That's so Jensen. Like, understands like, yeah, my employees get mad at me when I, you know, round the numbers and use CEO math. Like, I get it. Like, I appreciate precision too, but, you know, like, big picture is what matters here. The second way he falls in love is with his lab partner in, electrical engineering fundamentals.

他的实验室搭档洛莉后来成为了他的妻子。真是个精彩的故事。他1984年毕业，她1985年毕业。他们搬到了硅谷，黄仁勋加入AMD担任相当于芯片设计项目经理的职位。这个工作非常偏重工程，他有点像项目经理。

His lab partner Lori who goes on to become his wife. Such a cool story. So he graduates in 1984, she graduates in 1985. They moved down to Silicon Valley and Jensen joins AMD as a sort of equivalent of like a chip design PM. It's very like engineering heavy, he's kinda like a PM.

他有点像在协助担任芯片开发流程的初级经理。当时他正在研发一款速度惊人的1兆赫CPU芯片。

He's sort of like helping as a junior manager of of a process for developing a chip. He's working on a then blazing fast one megahertz CPU chip.

是的。他谈到这个时说，你知道，他讲到1兆赫有多慢，他形容说'你甚至能看到它慢慢过来'，就是形容当时的速度。

Yeah. He talks about this and he says, you know, he's talking about how slow one megahertz is and he refers to it and says, you could even see it coming. It's about how fast it was.

能看到它从

Could see it coming from

很远的地方过来

a long way away

还在慢慢过来，还在慢慢过来。太神奇了。而现在他制造的芯片确实是全世界最快的。他从AMD起步，晚上还在斯坦福攻读电子工程硕士学位。最终花了八年才完成这个硕士。

and still coming and still coming. Amazing. And of course, now he makes literally the fastest chips in the entire world. So he starts at AMD, he starts at night working on a master's degree in electrical engineering at Stanford. It ultimately takes him eight years to finish this master's.

他在AMD工作期间一直坚持学习，后来去了LSI Logic公司（我们稍后会谈到）。就在创立英伟达前不久，他终于毕业了。这是个非常有趣的冷知识。你有回去看过唐·瓦伦丁的《巅峰视角》吗？

He works all the time that he's at AMD and then at LSI Logic where he goes to. We're gonna talk about it in a sec. He ultimately does graduate right before they start NVIDIA. This is like a super cool bit of trivia. Did you go back and watch the Don Valentine, view from the top

不，我没有。

No. I didn't.

GSP的讲座。我每年都会看一次，年年如此。每次都能找到借口。

Lecture at GSP. I watch that, like, once a year, every year. Every time there's an excuse.

是不是他举着阿尔弗雷德简历的那次？

Is that the one where he holds up, Alfred's resume?

对，就是他举着阿尔弗雷德·林简历那次。那场演讲还有个彩蛋——那天正好是斯坦福大学Jensen和Lori Huang工程中心落成典礼。就像Dan说的，Jensen捐了栋楼。太厉害了。

It's yeah. Where he holds up Alfred Lin's resume. So also, Easter egg in that talk, that was the day that the Jensen and Lori Huang Engineering Center at Stanford was dedicated. And as Dan says, Jensen did a building. Pretty awesome.

我确实看过他在斯坦福的演讲，那是他走进来直接开讲。好像是Jensen自大楼启用后首次公开演讲。他说：'我已经捐过款了，现在有了这栋漂亮大楼，所以我现在身无分文了。'

I did watch he gives a talk where he he walks in and gives a talk at Stanford. I think it's the first time that Jensen has given a talk since the building opened. And he says, I've donated. We have this nice building now, so I I have no more money.

没错，我现在一贫如洗。

Yep. I'm penniless.

觉得是身无分文。对。

Think it's penniless. Right.

没错，Jensen。太棒了。

Right, Jensen. So great.

给大家提供个背景信息，如果你看看他的英伟达股票，他现在身价大约200亿美元。

Just to set context for people, if you look at his Nvidia shares, he's worth about $20,000,000,000 right now.

我想他持有英伟达大约3.5%的股份？差不多这样。他可不是穷光蛋。好吧。所以他在AMD工作了几年。

I think he owns what, like 3 and a half percent of Nvidia? Something like that. He's not penniless. Okay. So he works at AMD for a couple years.

他在那里工作期间，可能是在研发这款速度超快、肉眼可见的芯片时，意识到设计芯片真的太难了。英特尔能做到，AMD能做到。但你知道，没几家公司能做到。

And while he's working there, probably from working on this chip that you can so fast you can really see it coming. He realizes that designing chips is really freaking hard. Intel can do it. AMD can do it. But, you know, there's not many companies.

那时候都是全栈式的。台积电要到1987年才成立。

It's all, like, full stack at this time. You know, TSMC doesn't start till 1987.

不仅要在内部生产，而且在很大程度上，芯片设计过程是手工的。所以这些公司各自都有制度化的内部工作方式，用来设计和布局芯片元件。

Not only are you manufacturing in house, but for the most part, the, like, process of designing a chip is a manual one. And so these companies sort of each have their own institutionalized internal way of working that you design and lay out the elements of a chip.

Jensen谈到，他在学校时想去AMD的原因是他觉得这太酷了——你可以包办一切。但当他真的在AMD工作时，他意识到其实并不酷。如果能专注于某个环节，拥有工具和平台，并与其他公司合作让任何人都能制造芯片，那才更酷。

And Jensen talks about, like, when he was in school, the reason he wanted to go to AMD was he thought this was so cool that, like, you could do it all. And then once he's actually at AMD, he's like, he realizes, like, it's actually not cool. Like, it would be cooler if you could be really good at, like, a certain part of the stack and have tools and platforms and other companies to allow you to allow anybody to make chips.

是啊。如果有那种能帮你设计芯片的工具就好了。

Yeah. If there were, like, design tools to help you make chips.

几年后，他在AMD的同事离职加入了一家刚上市的初创公司LSI Logic。我们在节目里聊过，这让唐·瓦伦丁和红杉创投获得了当时史上最大的风险投资回报。可能是史上最大的风投回报——上市首日就达到了1.53亿美元。

So after a couple years, his office mate at AMD leaves and goes to join a startup called LSI Logic, which had just gone public. And we've talked about it on the show, made Don Valentine and Sequoia the then largest venture return in an IPO in history. Maybe the largest venture return ever in history when they went public of a $153,000,000 on day one.

天啊，风险投资这个资产类别真是今非昔比了。

Boy, has venture changed as an asset class.

但我试着回忆那个基金，可能是红杉二号或三号基金。我猜基金规模大概1.015亿美元左右，所以一天就赚了差不多整个基金的10倍。相当惊人。那么LSI是做什么的？

But that I'm trying to think that fund, that probably would have been, I don't know, Sequoia Fund two or three maybe. I mean, I bet the fund was like, I don't know, $1,015,000,000 like so probably roughly 10 x the fund in in one day. Right. Pretty awesome. So what was LSI?

它是最早也是顶尖的ASIC公司之一，ASIC即专用集成电路。他们主要为其他公司定制设计芯片，就像黄仁勋现在考虑的那样。这些定制的ASIC芯片会集成到其他系统中执行非常特定的功能。比如国防公司洛克希德·马丁等，现在越来越多公司都来找LSI Logic这类ASIC公司说：'嘿，我们想打造这种芯片系统'。

It was one of the first and was sort of the premier ASICs company, ASIC, Application Specific Integrated Circuit companies. And so what they did and what that meant was they basically made custom designed chips for other companies. It's what Jensen's kinda thinking about. And the custom designed chips that they would make these ASICs would be like for a very, very specific function that would be integrated into other systems. So like defense companies, Lockheed Martin, and the like, but lots of other companies now too are coming to LSA Logic and the other ASICs companies and saying, hey, we wanna create these systems of chips.

你们帮我们设计这些系统里的芯片，当然我们也会用英特尔处理器。但这确实让终端产品系统的制造变得更民主化了。

You help us design the chips to go into these systems and, yeah, we'll use processors from, you know, Intel too. But, like, it really helps democratize making end product systems.

没错。ASIC的理念就是：你不是在说'这需要驱动一台通用计算机，要超级灵活能运行各种应用'，而是'我知道这个芯片具体要做什么，而且它永远只做这个'。所以我们实际上可以直接把功能硬编码在芯片上。

Right. And the idea with ASICs is really if you're not saying, hey. There's gonna be a general purpose computer that this needs to power that can, you know, be super flexible and people might have all kinds of applications that run on it, but, you know, more inefficient in order to get that flexibility chip. Hey, I know the exact thing that this chip will do and it will only ever do this. And so we can actually literally hard code that right on the chip.

我是说，物理芯片的实际设计可以专门针对这一特定功能。所以它在处理这种底层任务时效率极高。

I mean, the the actual design of the physical chip can be for this one specific thing. So it's super efficient at this one low level thing.

没错。如今ASIC的遗产依然存在，两者都还在使用，但真正的遗产是FPGA（现场可编程门阵列芯片）。有些人可能会说这对英伟达来说是个不利因素，不过我们以后会详细讨论这个问题。Sun Microsystems曾是他们的最大客户之一，Sun就是这样起步并为工作站制造芯片的。实际上当黄仁勋加入LSI时，Sun才刚刚起步并开始与LSI合作，于是他被安排参与这个项目。在Sun微系统早期，他基本上就驻扎在Sun公司，帮助他们开发最终成为SparkStation一号的芯片——这是Sun首款重要工作站产品。

Yep. And the legacy of ASICs today, still around, still use both ASICs, but the legacy is FPGAs, field programmable array chips that are, you know, some might say sort of a bear case for NVIDIA these days, but we will we will get to that far far far down the road. Sun Microsystems was one of their biggest customers and that was how Sun got started and made the chips for their workstations. And in fact, Jensen, when he shows up at LSI, Sun is like just starting and coming to LSI and so he gets put on the project. He basically embeds with Sun, like in the early days of Sun Microsystems to help them build out the chips for what would ultimately become the SparkStation one, Sun's first big workstation product.

随后几年里，他在LSI Logic工作期间几乎只与Sun合作。他直接与Sun创始人Andy Bechtelstein以及Vinokoslub共事。是的，他因此声名鹊起，建立了能将Sun对芯片的构想和需求转化为实际产品的声誉。直到1992年感恩节前后，经过八年努力，黄仁勋终于完成了斯坦福大学的硕士学位。

Over the next few years, he pretty much exclusively works with Sun while he's at LSI Logic. He works directly with Andy Bechtelstein who, you know, the founder of Sun and with Vinokoslub. Yeah. He becomes super well known and develops quite a reputation there as somebody who can really like take these visions for chips and these customer requirements from Sun and turn it into, you know, reality and production. So one day, right around Thanksgiving nineteen ninety two, Jensen has finally, after eight years, finished his master's degree at Stanford.

斯坦福大学非常庆幸他在这件事发生前完成了学业。黄仁勋在Sun结识的两位好友Chris Malachowski和Curtis Pream——用他自己的话说，他们是极其出色的工程师。当黄仁勋这么说时，他是认真的。这两人找到黄仁勋说：'我们对Sun不太满意，有个想法想和你聊聊。'

And, Stanford is quite quite glad that he finished before this happens. Two of Jensen's buddies who he's become close with at Sun, Chris Malachowski and Curtis Pream, who in Jensen's own words, he describes them as really, really fantastic engineers. And when Jensen says that, he means it. They come to Jensen and they're like, we're not like super happy at Sutton, the two of us. We have an idea that we wanna talk to you about.

黄仁勋回答：'当然可以，去我最爱的丹尼斯餐厅谈吧。' 真的吗？没错，这人超爱丹尼斯。

And Jensen's like, well, sure. Let's go meet at my favorite spot, Denny's. Really? Yeah. Like, the man loves Denny's.

他高中时就在丹尼斯打工。至今还常去那儿，总点超级鸟套餐当招牌菜。太接地气了。

He worked at Denny's in high school. Like, he's always going to Denny's. He, he orders, the Superbird, I think, is like his go to dish. Nice. He's so folksy.

我太喜欢他了。三人在丹尼斯共进晚餐时，Chris和Curtis向他推销了他们的创意——这个点子确实不错。作为风投你会投资1992年底的这个创意吗？他们预见到3D图形技术即将崛起。

I love him. As they go all have dinner at Denny's and Chris and Curtis pitch him on their idea, which their idea is it's pretty good. It's pretty good. Tell me as a venture capitalist if you would fund this idea back then in late nineteen ninety two. So they see three d graphics are really becoming a thing.

要知道，那个时代属于Sun公司、LSI Logic等企业，同时也是硅谷的硅谷图形公司（SGI）的黄金时期。那里诞生了无数伟大事物——吉姆·克拉克、网景浏览器，还有即将上映的《侏罗纪公园》。

And, you know, remember this is the era of Sun, LSI Logic, all this stuff. It's also the era of Silicon Graphics right down the street, right there in Silicon Valley, SGI. So many great things that come out of there, you know, Jim Clark, Netscape, like all this great stuff. Jurassic Park. Jurassic Park is about to come out.

电影在1993年上映。当时市场对三维图形需求巨大，而制作三维图形必须依赖SGI工作站，需要极其专业、高端且昂贵的设备。只有军方预算或《侏罗纪公园》这种项目才负担得起。

It comes out in 1993. So there's this huge demand for three d graphics. The way three d graphics are done, you need SGI workstations. You need like super custom, you know, very high end, very expensive stuff. Only something with the budget of, like, either the military or, like, a Jurassic Park can afford to do this.

但人们为之疯狂，消费者对三维图形简直着了迷。

But people love it. Like, the consumers love three d graphics.

更不用说，当时游戏主机技术发展到什么阶段了？

Not to mention, where are we in the evolution of video game consoles at this point?

那时还处于超级任天堂时代，主机尚未实现三维图形，不过很快就会有突破。但个人电脑浪潮此刻正达到巅峰，

Well, we're still in the Super Nintendo days, so we're not at three d console graphics yet. That's coming very shortly. But what is happening is the PC wave is, like, really cresting right now. Like

距离Windows 95发布大约只剩一年半时间。

We're like a year and a half from Windows '95 coming out.

嗯。

Mhmm.

我也记得这件事。我打赌你也记得。1992、1993年的孩子们在电脑上做什么？他们在玩《德军总部3D》，哦还有《毁灭战士》。《毁灭战士》是1993年推出的。

And I remember doing this. I bet you do too. What are kids in 1992, 1993 doing on their PCs? They're playing Wolfenstein three d Oh, and Doom. Doom comes out in 1993.

这些游戏正风靡全球，它们由id Software（德州的公司）的约翰·卡马克和约翰·罗梅罗开发。但卡马克当时正完成一项工程壮举——让3D图形能在消费级PC上运行。需要卡马克这种水平的天才才能实现这个突破，而市场反响极其热烈。所以克里斯和柯蒂斯的想法是：'我们是顶尖的芯片工程师，黄仁勋你本质上是个出色的芯片产品经理'。

These are taking the world by storm and they're made by id, id Software in Texas and John Carmack and John Romero. But Carmack is like doing incredible feats of engineering to get three d graphics to run on consumer PCs. It took somebody of Carmack's caliber to make this happen and the market loved it. So the idea that Chris and Curtis has, they're like, we're really great chip engineers. Jensen, you're a really great, you know, chip PM essentially.

我们来做显卡吧！开发一款能加速普通PC图形处理的芯片，让消费级硬件也能实现像SGI专业工作站那样的3D图形。我们知道人们热爱游戏，这将推动整个行业腾飞。

Let's make a graphics card. Let's make a chip that can accelerate the graphics of a normal PC to enable three d graphics like SGI is doing with professional workstations to enable them for consumer hardware PCs. We know that people love games. This will help the entire industry, you know, take off.

你甚至都没提到

And you're not even saying

听起来不错吧？

Sounds pretty good. Right?

他们不是说要让游戏能在PC上开发，而是说让游戏能在PC上玩对吧？

That they're gonna try and make it so you can develop games on a PC. You're saying, like, just so you can play games on a on a PC. Right?

两者都有。主要是让玩家能在PC上玩游戏，但同时也需要建立API、SDK和开发者生态，让开发者能调用PC的新硬件——不过他们直接在PC上开发就行。核心是要让消费者能买到硬件来运行这些游戏。你觉得呢？

Well, both. I mean, mostly that you can play games on the PC, but then you're also gonna have to create, you know, all the APIs and SDKs and developer ecosystem for developers to access this new hardware on PCs, but they'll just develop on PCs. It's really about getting the like, the hardware into consumers hands that they can actually play this stuff. Alright. So what do you think?

这个提案听起来不错吗？

Is this like a good pitch?

我是说，你基本上是在让我（1992年的我）相信，PC上的电子游戏将成为一个大趋势，会掀起经济浪潮，大量消费者会热衷于此。他们会选择在PC上玩，而不是超级任天堂或专用游戏机。也许吧。但我有个证据

I mean, so what you're basically asking me to believe, 1992 me, is that video games on PCs are gonna be a thing that there's gonna be a big economic wave around, that lots of consumers are going to want to do this. They're gonna wanna do it on PCs instead of on Super Nintendo and dedicated systems. Maybe. Well, I have this proof point

就是id Software公司以及《德军总部》和《毁灭战士》的存在。看，已经有数百万人这么做了。

of of id Software and and Wolfenstein and Doom right there. Like, millions of people doing this.

但也许还是因为，电子游戏是否会成为巨大市场尚不明确。可能只是儿童市场，你知道吗？而且问题是，真的需要彻底改变开发环境吗？或者说，能不能同时存在五六个不同的《毁灭战士》？有五六个卡马克这样的天才各自独立地解决图形技术问题。

But still maybe because it's not clear that, like, video games are gonna be a giant market. It could be like a kid market, you know, and it could be the case that, like, do you really need to totally change the development environment? Or can, like, there be, like, five or six different Dooms out there? There's five or six Carmacks who are all independently geniuses and can figure out how to do all the graphics on their own. Yeah.

也许吧。但这需要信念的飞跃。

Maybe. But there's a leap of faith.

是的。确实需要信念的飞跃。所以没错，这并不完全显而易见。

Yeah. Definitely a leap of faith. So yeah. Okay. Not totally obvious.

但我仍然认为，在当时这个时间点，这个项目是相当值得投资的。同时硅谷还有周边产业在兴起，比如制造插入消费者PC的芯片和扩展卡的公司。这正处于全盛期——有公司做声卡，有公司做网卡。

But still, like, I think this was pretty fundable, I think, at this moment in time. And the other thing that was going on was in Silicon Valley of these peripheral companies, like people building chips and cards that plug into consumers PCs. This was full swing. There are companies making sound cards. There are companies making networking cards.

有些公司在生产串口卡之类的玩意儿，天知道是些什么东西。

There are companies making serial port cards, like god knows what.

好的。现在已经有股加速计算的浪潮，人们认为有必要将某些专门任务从CPU中分离出来，需要专门的集成电路，厂商正在定制这类产品，存在一个为PC制造定制硬件的市场，可以分担CPU的工作负载。

Okay. So there's already, like, sort of an accelerated computing wave here where people are saying, like, there's some reason to do something specialized off the CPU that needs its own integrated circuit that vendors are making custom, and there's a market to make custom stuff as a vendor for PCs that takes a workload off the CPU.

没错。我们的卖点是要制造一款定制显卡，专门为游戏分担图形处理工作负载。很好，明白了。

Yep. And so the pitch is we're gonna make a custom graphics card. Take a graphics workload off the CPU, specifically for gaming. Great. Okay.

所以当时这简直是毋庸置疑的选择。但正如节目开头提到的，当某件事对风投来说显而易见时，问题就在于——所有风投都这么想，于是会有无数公司获得资金来做这件事。但回到那晚的Denny's餐厅，英伟达是首家专注显卡的公司。他们三人决定共同投身于此。

So, yeah, it was pretty much a brain dead yes. But as you alluded to at the top of the show, the problem when something is a brain dead yes for venture capitalists is that it's a brain dead yes for lots of venture capitalists and lots and lots and lots of companies get funded to do this. But back to Denny's that night, NVIDIA is the first. They are the first dedicated graphics card company. They all decide, the three of them, that they're gonna go in on this.

黄仁勋去找LSI Logic的CEO，走进办公室告诉对方自己要辞职。他将和两位来自Sun的工程师创业，这就是商业计划的内容。你们知道当时LSI Logic的CEO是谁吗？是威尔夫·科里根，他之前是仙童半导体的CEO。

Jensen goes to the CEO of LSI Logic, walks into his office and tells him that he's gonna resign. He's gonna go start this company with two engineers from Sun, and, this is what the business plan is gonna be. Now do you know who the CEO of LSI Logic was? No. It was a man named Wilf Corrigan who was previously the CEO of Fairchild Semiconductor.

没错。太对了。

Way. Damn right.

所以唐就是这样介入的？因为唐·瓦伦丁显然是LSI Logic的最大投资者，或者说红杉资本投资了LSI Logic，他是在仙童时期就认识科里根的吗？

So is that how Don because Don Valentine obviously was the biggest investor in or Sequoia was in LSI Logic, and he did he know him from Fairchild?

是的。他们当年是同事。好吧。然后创下了红杉资本当时历史上最大的一笔退出。于是威尔夫说，让我捋清楚。

Yeah. They were colleagues back in the day. Okay. And then the biggest exit in, in Sequoia's history to that point in time. So Wilf says, so let me get this straight.

他对黄仁勋说，你去造这些显卡，就像你刚才说的那样，本。谁会用这些？用来干什么？就像，嗯，你知道的，它们会用在个人电脑上。用于游戏。给一群孩子用的。

He says to Jensen, you're gonna go build these graphics cards and kind of just like you were saying there, Ben. Who's gonna use these and and what for? It's like, well, you know, they're gonna be in PCs. They're for gaming. They're for a bunch of kids.

而沃尔夫抓住了关键问题。他说，那么，谁开发PC游戏？这个有开发者生态吗？这有点像，我们觉得只要造出来，就会有人用。沃尔夫说，记得他是个公正的孩子。

And Wolf hones in on the critical question. He's like, well, who makes PC games? Is there a developer ecosystem for this? So that's kinda like, we think if we build it, like, they'll come. So Wolf says, remember he was a fair child.

他觉得自己知道什么时候该为特定应用制造芯片。然后沃尔夫说，好吧。你会回来的。我会给你留着位置。但在此之前，在你走之前，我要给唐打个电话。

He felt like he knows when to make silicon for specific applications. And, Wolf says, alright. You'll be back. I'm gonna hold your desk. But in the meantime, before you go, I'm gonna call up Don.

我要说，你为我干得不错。我要给唐打电话。他打给唐，说，唐，我这儿有个小伙子。他要来见你。等着吧。

I'm gonna do you've done good work for me. I'm gonna call up Don. He calls up Don and he's like, Don, I got a kid. He's gonna come see you. Stand by.

这对所有创始人和有志创业者都是个教训。获得投资组合公司CEO的推荐，是让风投倾向于投资的好方法，尤其当你被他们投资史上表现最佳的公司推荐时。没错。带着这样的推荐，黄仁勋几乎不可能搞砸这次推介。

Which this is a lesson for all founders and, you know, aspiring founders out there. Getting a reference from the CEO of a portfolio company is a really good way to come in with a venture capitalist already leaning toward investing, especially if you're referred by the top performing company of all time in their portfolio. Yes. It's kind of hard for Jensen to mess up this pitch with the recommendation that he's coming in with.

这简直是不可能的，因为他去见唐。你知道的，唐。唐坐下来就问，所以呢？结果黄仁勋把推介完全搞砸了。他变得非常紧张。

It's literally impossible because he goes to see Don. You know, Don. You know, Don sits down and he's like, so? And Jensen completely botches the pitch. He gets, like, really nervous.

当时我觉得他大概写了份半成品的商业计划书，买了本创业指南之类的书，看了三章就放弃了，然后开始写计划书又没写完。所以他来开会时，就像对着唐一股脑儿全倒出来了。

At this point, I think he had, like, a partially written business plan that he had, like, bought a book on, like, how to start a business and was, like, three chapters into the book, but decided not to finish and started writing the plan and didn't finish the plan. So he comes into this meeting and just kind of, like, barfs all over Don.

没错。就是这样。詹森垂头丧气地正要离开，唐拦住他说，嗯，刚才表现可不怎么样。

Yes. Exactly. So Jensen's walking out the door. He's, like, you know, totally dejected. Don stops him and says, well, that wasn't very good.

但是威尔夫说要给你钱。所以尽管根据你刚才说的内容，我完全不该投钱，但我还是决定给你投资。不过要是你把我的钱赔光了，我就宰了你。典型的唐式台词。

But, Wilf says to give you money. So against my best judgment, based on what you just told me, I'm gonna give you money. But if you lose my money, I'll kill you. Classic. Classic Don line.

太精彩了。于是交易达成。萨特山风投也加入了。因为你知道，艾丽莎把这事渲染得戏剧性十足。说到底这是个热门项目。

So good. So the deal happens. Sutter Hill comes in too. Because, know, again, like, Alyssa's all dramatized. At the end of the day, like, this is a hot deal.

这是我们连续第二期节目提到萨特山了。

This is, two episodes in a row for us with, Sutter Hill.

我知道。天啊。他们太厉害了。但那确实是个香饽饽，大家都抢着要。

I know. Oh, jeez. They're so good. But it was a hot deal. They wanted it.

这完全符合当下这个时间点的典型剧本。

This fits central casting of at this point in time.

他们每人投资了大约一百万，对吗？总共两百万？

They invested, like, a million each. Is that right? For a total of two?

所以这轮总共200万美元。我不知道具体谁投了多少。我猜每人一百万，但总额是200万美元，投后估值600万美元。记住各位，这可是当时全球第八大最有价值的公司。起步时的投后估值是600万美元。

So $2,000,000 total round. I don't know who invested what. I assume a million each, but $2,000,000 total round at a $6,000,000 post money valuation. Remember, everybody, this is the eighth most valuable company in the world right now. Started at a $6,000,000 post money valuation.

所以他们正在解决各种问题，但还有一个问题——公司还没起名字。黄仁勋、克里斯和柯蒂斯一直在忙这个，忙着做商业计划，但还没定下名字。他们需要注册公司，当时保存首个图形芯片设计文件的扩展名是.nv，nv代表next version（下一个版本）。然后他们觉得，哎，我们还挺喜欢这个的。

So they're getting things ironed out and there's just one problem, they don't have a name for the company yet. Jensen and Chris and Curtis, they've just been, you know, working on this, working on the business plan, they don't have a name. They need to incorporate the company and they were saving the files that they were working on for the chip design for the first graphics chip as dot n v. N v being short for next version. And so like, oh, we kinda like that.

你看，我们总是在开发下一个版本嘛。他们开始查字典找带NV的词。可能列表很短，结果找到了拉丁词Invidia（拼写为I n v I d I a），意思是嫉妒。他们觉得太棒了，我们要成为行业嫉妒的对象。

You know, we're always working on the next version here. They start looking around in the dictionary for words that have NV in them. It's probably a very short list and they find the Latin word Invidia, I n v I d I a, which means envy. And they're like, great. We'll be the envy of the industry.

NVIDIA会去掉开头的I，这样就从envy（嫉妒）开始了。这太酷了。

NVIDIA will drop the I at the beginning, so we start with envy. This is awesome.

当然，他们选了绿色。这样以后就能做'嫉妒得发绿'的营销活动了。

Of course, they pick green. So later on, they can have that marketing campaign of green with envy.

不过在这里要小心你的愿望，因为正如我们所说，短短几个月内就有其他89家公司获得融资来做同样的事情。

Careful what you wish for here though because, again, as we've been saying, 89 other companies get funded within a couple months to go do the same thing.

这是个非常聪明的名字。而且，'vid'这个词根暗示了视频元素，这正是他们想做的另一件事。就像经典的Rich Barton式空容器命名法——这个名字有足够的可解释空间，他们会赋予它意义。因为他们正在做的事情，实际上还有89个其他人也在同步进行。

It's a very clever name. Also, the notion of, like, vid being in there that it's sort of video, and that that's another thing that they wanna do. Like, it's the classic Rich Barton empty vessel name. You know, there's enough things that it could mean, and we're gonna fill it with with meaning. Because they're doing a thing here that, like, well, 89 other people are also sort of simultaneously doing.

这确实是一个需要他们开创并主导思想领导权的新领域。他们需要快速建立品牌，不仅面向消费者，还要面向PC制造商。黄仁勋描述他们的愿景时（虽然他不喜欢'愿景'这个词，认为它有排他性）说：'我们的观点是希望让图形技术成为叙事的新媒介。'他当时这样阐述：如今电子游戏已成为年产值1800亿美元的产业，规模超过好莱坞和音乐产业。

It is kind of a new frontier that they need to invent and then own, like, thought leadership in that area. And they do need to kind of, like, quickly build a brand not only with consumers, but with PC manufacturers. Jensen, the way he sort of describes it is that their vision although he doesn't like the word vision because he thinks it's exclusionary to people. So he said our perspective is that they wanna enable graphics to be a new medium to tell stories. And here's sort of like the way that he articulates at the time why video games today are a $180,000,000,000 a year industry, bigger than Hollywood, bigger than music.

这是最大的娱乐媒介。但当时他有个观点：通过计算机图形技术，你其实还无法真正实现叙事。但如果能做到就非常有趣，因为它不是预录制的，每次体验都能有新的变化。这也是唯一可以网络化的娱乐媒介，因此也是唯一真正具有社交性和互动性的形式。

It's the biggest entertainment medium. But at the time, he sort of has this thesis that, like, you really can't, through computer graphics, tell stories today. But if you could, it's really interesting because it's not prerecorded, so it can be sort of new and different every single time you enjoy it. It's also the only medium of entertainment that can be networked. And so therefore, it's the only one that can really be like social and interactive.

所以我们存在的意义，就是为未来创造作为艺术叙事形式的三维图形，一切都将服务于这个目标。我认为这并不完全是他们现在的全部，但正是这个理念支撑了他们最初二十年的发展。

And so our reason for being is to create three d graphics as a form of artistic storytelling for the future, and everything will be in service of that. And I think that's not really what they are today necessarily. It's a piece of what they are today, but that kept them going for the first twenty years of their existence.

这里还隐含着一个问题，Wolf刚才也提到了，你也说得很到位——你是个非常优秀的风险投资人。你指出了英伟达初代产品的关键问题：他们必须向开发者布道。当时除了id和Car Mac外，PC游戏开发者本就不多，三维PC游戏开发者更是凤毛麟角。

Well, and baked into that is again, you know, Wolf kinda like hit on it. And you did too to your credit. You're a very good venture capitalist. You hit on really the key problem with this first iteration of Nvidia, which is they have to go evangelize to developers to, like, yeah, there's id and there's car mac out there, but, like, not a whole lot of other PC game developers out there. There's not a whole lot of other three d PC game developers at this time.

虽然有些2D PC游戏开发者，但他们需要说服大批人去学习PC端3D游戏开发。就像在说'我们要赋能所有叙事可能性'。为此他们必须自己编写API、SDK和开发框架，为这款新显卡建立开发生态，还要做出一系列希望行业能标准化采用的技术设计决策。

There are two d PC game developers, but they gotta convince a whole lot of people to go, you know, learn how to do three d game development for PCs. And that's so like, oh, we're gonna enable storytelling all them. So to do that, they have to go write their own, you know, APIs and SDK and development framework to develop for this new graphics chip that they come out and they have to make a whole bunch of, like, technical design decisions that they want the industry to standardize on.

没错。这就是当你比整个行业都超前时会发生什么的典型案例。

Right. This is a case study of what happens when you get more clever than the rest of the industry.

没错。一开始，事情进展得非常顺利。记住，这可是炙手可热的项目。他们是首家获得红杉资本和萨特山投资的公司。

Exactly. So at first, things start off really well. Remember, this is super hot. They're the first company. They're funded by Sequoia and Sutter Hill.

比如，他们与世嘉达成重大合作，为其街机游戏机和下一代家用主机提供3D图形引擎支持，最终成为世嘉土星主机。正如我们从索尼那期节目中所知

Like, they land a big deal with Sega to power their arcade consoles and their next generation home console to be the three d graphics engine and would ultimately become the Sega Saturn. And as we know from our Sony episode

并不完全是世嘉创世纪主机。

Not quite the Sega Genesis.

确实不完全是世嘉创世纪。问题是，英伟达和世嘉当时正在合作。他们做了一系列设计决策，其中最重要的是决定采用多边形技术——大家都知道3D图形是用多边形构建的，这就是为什么这个行业总在讨论多边形。他们必须确定多边形的基本单元。

Not quite the Sega Genesis. Well, so the problem is, so NVIDIA and Sega, they're working together. They make a bunch of these design decisions, the biggest of which is they decide that the way they're gonna create you know, people probably know you create three d graphics, use polygons. That's why people are always talking about polygons in this industry. They have to decide on a sort of primitive for the polygon.

他们当时决定：'哦，我们用四边形作为顶点吧'。现在任何懂游戏开发的人都会说：'根本不是这样做的'。

They're like, oh, well, we'll use quadrilaterals for vertex. You know? And anybody who knows anything about video game development now is like, that's not how it's done.

对，我记得人们讨论的都是三角形。

Like, I'm pretty sure people talk about triangles.

没错。而且我很确定，如果你现在看英伟达令人惊叹的总部大楼，就会发现它是由三角形构成的——这是向游戏开发者致敬，而不是四边形。所以这成了个大问题。事情持续了一段时间，大约一年内还算顺利。

Yeah. And I'm pretty sure if you look at NVIDIA's amazing headquarters building today, it's, you know, made out of triangles in a homage to game developers, not quadrilaterals. So this becomes a pretty big problem. You know, things go along for a while. It's like fine for about a year.

英伟达处于领先地位。他们拿下了世嘉这个大单。

NVIDIA's leading. They got this big Sega deal.

目前还没有需要制定标准的理由，对吧？这个行业还不够复杂，尚不需要大量协作和一套大家都标准化使用的工具。你会觉得，好吧，我们就把这个芯片放进游戏主机，然后发售就行了。

There's not a reason to need standards yet. Right? The industry isn't complex enough yet to necessitate a whole bunch of collaboration and set of tools that everyone standardizes on using. You're like, okay. Well, we're just gonna put this chip in our game console, ship the game console.

我们是唯一开发SDK的，这里'我们'指的是世嘉。所以无论如何大家都得在一定程度上以这个为标准。这很好。但显然，生态系统会迅速变得复杂得多，如果能有些兼容性就太好了。

We're the only people that, you know, make an SDK, we being Sega. So everyone will have to kinda standardize on this thing anyway. So great. But obviously, the ecosystem gets much more complex, much more quickly, and it sure would be nice to have some kind of compatibility.

事情是这样的。柯蒂斯、克里斯和黄仁勋并不是硅谷唯一注意到孩子们想在PC上用《毁灭战士》玩游戏的人。微软的反应是，哦，这很有趣。我们喜欢卖PC，天哪，现在有这么多显卡公司在做这个，我们微软该怎么办？我们真的很想在生态系统中鼓励这种现象。

Well, here's what happens. So, you know, Curtis and Chris and Jensen, they weren't the only people in Silicon Valley that saw that kids wanna play games on PCs with Doom. Microsoft is like, oh, that's interesting. We like selling PCs and, gosh, there are all these graphics cards companies out there now that are doing this and know, what do we do as Microsoft? We really wanna encourage this in the ecosystem.

于是，我们创建标准。

Well, we create standards.

如果Windows开发者能轻松地为这些搭载先进图形功能的新机器开发软件就太好了。让我们尽可能简化开发者的工作。

We would love it if Windows developers could be able to easily develop for all these new machines shipping with all these advanced graphics capabilities. Let's make that as easy as possible for those developers.

没错。开发者想直接在Windows上做3D图形，不需要那些不知名公司（比如英伟达）的狡猾中间件。我们何不把这些API直接集成到Windows里做3D图形？就叫它Direct3D吧。当然，了解这段历史的人都知道，后来这就发展成了DirectX。

Yeah. You know, developers wanna do three d graphics directly into Windows without any of this, you know, crafty middleware from some no name company, you know, NVIDIA out there. Why don't we just bake these APIs right into Windows directly for three d graphics? We'll call it Direct three d. And, of course, anybody who knows about the history of this, that becomes DirectX.

DirectX做出的设计决策与NVIDIA大相径庭，是这样吗？

And DirectX made some pretty different design decisions than NVIDIA had made. Is that right?

没错。他们选用三角形因为确实合理。现在NVIDIA就真陷入困境了。要知道其他89家后来入局的竞争对手，大多都想着‘当然要加入微软生态，不跟才是傻子’。

Yeah. So they use triangles because triangles make sense. So now NVIDIA's really up a creek. Like, of their you know, the 89 other competitors out there that started later, most of them are like, sure, I'm gonna jump on board to this Microsoft ecosystem. Like, I would be dumb not to.

这个标准完全背离了Nvidia和世嘉的范式。他们当时只有世嘉这个客户。结果1996年世嘉也表示‘我们对四边形方案也不太看好’。

It's standardized on this completely different paradigm than Nvidia and then Sega. You know, they've got Sega. They've got this one sort of customer. And then in 1996, Sega's like, yeah. We're not so sure about this quadrilaterals thing either.

为了避免讨论显得突兀，我们在此仅从宏观层面简述3D图形原理。三角形是构成二维形状所需的最少顶点数，因此成为基础构建单元——只要绘制足够多且足够小的三角形，就能组合出任何其他形状或曲面，它是创造3D视觉效果最基础的构件。

And just so that, like, this doesn't feel arbitrary why we're talking about this. And we're gonna say at a super high level on three d graphics here rather than really going into the weeds. But a triangle is the fewest vertices in a shape that you can have while still creating a two dimensional shape. And so it serves as a basic building block where assuming you can draw enough triangles and make the triangles small enough, you can form any other shape, any other curved surface. It's sort of the most fundamental building block that you could use to create something that is perceived as three d.

对。

Yep.

好了听众朋友们，现在聊聊我们最爱的公司之一Statsig。自上次报道后他们有个激动人心的进展：C轮融资后估值达到11亿美元。

Alright, listeners. It's time to talk about another one of our favorite companies, Statsig. Since you last heard from us about Statsig, they have a very exciting update. They raised their series c, valuing them at $1,100,000,000.

确实是重大里程碑，祝贺团队。时机也很有趣，因为实验领域现在竞争正日趋激烈。

Yeah. Huge milestone. Congrats to the team. And timing is interesting because the experimentation space is, really heating up.

是的。那么为什么投资者对STAT SEG的估值超过十亿美元？因为实验已成为全球顶尖产品团队产品栈中的关键部分。

Yes. So why do investors value STAT SEG at over a billion dollars? It's because experimentation has become a critical part of the product stack for the world's best product teams.

没错。这一趋势始于Facebook、Netflix和Airbnb等Web2.0公司。这些公司面临一个问题：如何在员工规模扩大到数千人的同时，保持快速、去中心化的产品和工程文化？实验系统正是解决方案的重要组成部分。

Yep. This trend started with web2.0 companies like Facebook and Netflix and Airbnb. Those companies faced a problem. How do you maintain a fast, decentralized product and engineering culture while also scaling up to thousands of employees? Experimentation systems were a huge part of that answer.

这些系统让公司每个人都能获取全局产品指标，从页面浏览量到观看时长再到性能表现。每当团队发布新功能或产品时，他们都能衡量该功能对这些指标的影响。

These systems gave everyone at those companies access to a global set of product metrics, from page views to watch time to performance. And then every time a team released a new feature or product, they could measure the impact of that feature on those metrics.

因此Facebook可以设定'增加应用使用时长'这样的公司级目标，让各团队自行探索实现方法。当数千名工程师和产品经理同时推进时，就会产生指数级增长。难怪实验现在被视为必要的基础设施。

So Facebook could set a company wide goal like increasing time in app and let individual teams go and figure out how to achieve it. Multiply this across thousands of engineers and PMs and boom, you get exponential growth. It's no wonder that experimentation is now seen as essential infrastructure.

是的。如今Notion、OpenAI、Rippling和Figma等顶级产品团队同样依赖实验。但他们不再自建系统，而是直接使用Statsig。而且Statsig不仅用于实验——过去几年它还添加了快速产品团队所需的所有工具，如功能开关、产品分析、会话回放等。

Yep. Today's best product teams like Notion, OpenAI, Rippling, and Figma are equally reliant on experimentation. But instead of building it in house, they just use Statsig. And they don't just use Statsig for experimentation. Over the last few years, Statsig has added all the tools that fast product teams need like feature flags, product analytics, session replays, and more.

所以如果你想帮助团队的工程师和产品经理更快构建并做出更明智的决策，请访问statsig.com/acquired，或点击节目说明中的链接。他们提供非常慷慨的免费套餐、5万美元的初创企业计划，以及适合大企业的实惠合同。只要告诉他们是本和大卫推荐你的。

So if you would like to help your team's engineers and PMs figure out how to build faster and make smarter decisions, go to statsig.com/acquired, or click the link in the show notes. They have a super generous free tier, a $50,000 startup program, and affordable enterprise contracts for large companies. Just tell them that Ben and David sent you.

当时英伟达已经完成了他们认为世嘉会采用的下一代芯片（最终成为Dreamcast）的一半开发工作，他们称之为NV2。当世嘉突然表示要更换方案时，英伟达就陷入了多重困境。除了我们讨论过的所有问题，在这期间——自英伟达启动项目后的一年半里，内存价格还因摩尔定律下跌了。

So NVIDIA at this point, they're halfway down the road of developing the next chip that they think Sega is gonna adopt for what ultimately would become the Dreamcast. NVIDIA was calling the NV two. When Sega comes back and says, we're switching horses, we're not gonna do this. So, like, they're screwed for so many reasons. Everything we've discussed, there's also in the interim, you know, year and a half since NVIDIA started, the price of memory dropped because thank you Moore's law.

所以英伟达的芯片设计对内存极其苛刻，内存组件成本约200美元，你知道的，就是芯片里的那些部件。而他们的竞争对手拥有更多内存，成本却只要50美元左右。

So NVIDIA's chips were designed to be like super super tight on memory and the memory cost about $200 in component, you know, parts to go into their chips. Their competitors have more memory that's costing them, like, $50.

而这仅仅是那一代产品的情况。有趣的是，英伟达作为先行者，由于没有预见到摩尔定律带来的指数级变化，实际上处于劣势。首先，他们没能观察行业标准如何确立，结果自己选了一条技术路线，最终与主流背道而驰——这反而成了优势。但其次，其他厂商的成本结构低得多，或者说至少能预见成本将大幅下降。

And that was just in that one iteration. So it's interesting to note that NVIDIA by being first and not projecting out the exponential change that would come from Moore's Law was actually at a disadvantage. Because a, they didn't get a chance to watch and see where the standards were adopted. And so they sort of like picked their own lane and went off in their own direction, which ended up not being what everyone else picked, which put them at its advantage. But second of all, everyone else's cost structure was way lower, or at least everyone else could see that the cost structure was getting way lower.

因此英伟达的设计基于一个已失效的前提——当竞品面世时，这个约束条件已不存在。这时黄仁勋和联合创始人不得不面面相觑：我们要全盘推翻现有设计吗？如果要，如何避免重蹈覆辙？如何确保未来能预判摩尔定律的指数曲线和价格下降趋势，为两三代之后的技术需求做设计？

And so NVIDIA sort of designed for a constraint that was no longer true by the time everyone else came out with their stuff. At this point, Jensen and his cofounders kind of had to look at each other and say, okay. Do we scrap everything we did? And if so, how do we not make this mistake again? How do we make sure that in future generations, we sort of premeditate the exponential curve of Moore's law and prices coming down and design for things that are, you know, two, three, four generations beyond what we actually have available to hardware right now?

当这一切发生时，公司只剩9个月周转资金。换作任何人都会选择放弃——游戏结束了，所有牌面都对你不利，彻底完蛋了。

So when all this goes down, the company has about nine months of runway left. And, like like, literally anybody else, like, you pull the plug. Like, it's over. Like, everything in the deck is stacked against you. Like, you're effed.

我无法想象当时他们如何构思破局之道。但老黄啊，他真是个狠角色。他说：我们绝不能这样认输。如今你听黄仁勋谈英伟达文化时，他说'求真精神是英伟达文化的基石'——

And, I can't imagine sitting there dreaming up a way out of this. But Jensen, god, he's such a g. He's like, no. We're not going out like this. You know, when you hear Jensen talk today about like NVIDIA's culture and he says that intellectual honesty is like the cornerstone of NVIDIA's culture.

这就是他当年在践行的事。他和Curtis、Chris开会时——别忘了这些人都是工程师，当时他们已为英伟达招募了100多名工程师，并向他们兜售'我们要定义行业标准，绝不用现成方案'的技术愿景——

Like, this is what he's freaking talking about. Like, he sits down with Curtis and Chris. And remember, they're like they're engineers and they've recruited NVIDIA a 100 plus engineers into the company at this point and sold them on this technological vision of we're gonna define the industry. We set the standards. Like, we're not gonna use some, you know, off the shelf stuff and like it's all toast.

于是黄仁勋说：兄弟们，这根本是痴人说梦。要想活下去就得推倒重来。我们唯一能做的就是采用和所有人一样的微软Direct3D标准、相同架构，唯一机会就是在性能上竞争，在这片同质化芯片的海洋中做出最好的产品。而他的联合创始人表示：这太无趣了，根本不是硅谷工程师向往的愿景。

And so Jensen's like, guys, like this is a pipe dream. We need to throw it all out if we're gonna survive. The only thing we can do is standardize on on the same Microsoft, you know, Direct three d as everyone else, same architecture, and our only shot is just to, like, compete on performance and try and become, like, the best chip out there in this now sea of commodity chips. And, you know, his cofounder is, like, don't wanna do this. This is not an exciting vision for a Silicon Valley engineer.

当你的CEO来找你并这么说时，他们本质上是在表达：听着，如果我的职责是制定战略而你的任务是执行，那么战略已经失败了。所以我们现在需要实实在在地在工程能力上超越所有竞争对手。我们必须做出更明智的工程决策，从而以更低的价格、更少的能耗实现比竞争对手更优的性能。因为微软作为微软，已经吸引了所有开发者的注意力。由于微软制定了标准，英伟达意识到：看，我们当时根本没有能力独树一帜地吸引自己的开发者群体。

When your CEO comes to you and says that, what they're basically saying is, look, if my job was strategy and your job is execution, the strategy failed. And so we just now need to, like, literally out engineer all of our competitors. We need to be smarter at engineering decisions so we can be more performant at a lower price point using less energy than our competitors. Because Microsoft, being Microsoft, had all the developer attention. And because Microsoft set a standard, NVIDIA realized, look, we have no ability to uniquely get our own developers, at least at that point in the company's history.

因此我们只能眼睁睁看着左边——所有开发者都通过微软的API涌来；右边是同样的消费群体，我们不得不在纯粹的工程能力上与所有人正面竞争。

And so we must just, on our left, look and see all the developers are coming from Microsoft using this API. On our right is all the same consumers, and we have to compete just head to head on raw engineering ability with everyone else.

你提到工程能力。但记住，这本质上已经是同质化商品了。所以关键不仅是工程能力，更是交付速度——你能多快设计出下一代芯片并抢先上市？因为大家都知道芯片里会有什么。

Well, you're saying engineering ability. But remember, like this is essentially a commodity at this point. So really it's not just engineering ability, it's how fast can you ship. Like how fast can you design the next generation of chip and can you ship it before everybody else? Because everybody knows what's gonna be in that chip.

为什么会这样？显卡究竟有什么根本特性使其沦为同质化商品？

And why is it? What fundamentally about was it about graphics cards that made it a commodity?

到这个时候，就像其他外设一样（我们稍后会详谈），它并没有什么特别之处。它们都在做同样的事——将多边形级别的3D图形处理从CPU转移到主板上的另一块芯片，就像声卡处理声音、网卡处理网络那样。问题只在于：这么做的性价比如何？接口和编程语言都已被微软标准化了。

Well, at this point, like all the other peripherals, and we're gonna get into this in a sec, there was nothing that special about it. They all did the same thing, which was take polygon level three d graphics processing out of the CPU and onto this other chip on the motherboard. Just like sound cards were doing the same thing for sound, just like networking cards were doing the same thing for networking. And it was just like, what's the price performance ratio of doing that? The interfaces and the programming language, that's all standardized by Microsoft.

你们只是同质化硬件罢了。

You're just commodity hardware.

所以GPU实际做的（至少在当时）就是：系统会给我输入点云数据，也就是构成3D世界多边形的顶点。而我的职责就是尽可能快速、以尽可能高的分辨率（或者说预设的标准分辨率）...

And so what GPUs actually do or did at least in this point in time is say, okay, the system is gonna feed me in basically point clouds, like vertexes that make polygons that represent like a three d world. And my job as the GPU is to, as fast as I can in the highest resolution that I can, or I suppose a standard predetermined resolution

尽我所能地快。这将决定最终的分辨率。

As fast as I can. And that'll drive the resolution.

输出一个二维的东西到屏幕上。所以我把三维的东西转换成二维的，而且我必须做得比其他竞争对手更好，基本上我们所有人都在竞争。当你说商品化时，指的是受摩尔定律限制，并且要一直推进到集成电路制造技术允许的极限。

Output a two d thing that goes on the screen. So I turn three d stuff into two d stuff, and I have to do that better than other things that I'm competing against where basically all of us are. When When you say commodity, you mean limited by Moore's law and doing right up to the edge of what integrated circuit manufacturing techniques enable us to do.

没错。所以大家都知道这意味着他们必须比竞争对手更快推出产品。他们还必须比竞争对手更快，因为他们即将破产。所以他们制定了这个计划，就像在尝试穿过最细的针眼。他们必须裁掉公司70%的员工，他们也确实这么做了。

Yep. So everybody knows what this means is that, like, they gotta ship faster than their competitors. They also gotta ship faster than their competitors because they're about to go bankrupt. So they draw off this plan that's like they're trying to thread like the tightest needle possible here. They have to lay off 70% of the company which they do.

他们缩减到大约35人，留下的人都知道我们现在必须从零开始设计，并在资金耗尽前（九个月内）推出新芯片。这在正常的芯片设计周期内是不可能完成的。

They go down to about 35 people And everybody who's staying knows we now have to design from scratch and ship a new chip before our runway runs out, which is nine months. You can't do that on a normal chip design cycle.

通常需要两年左右。对吧？

Takes like two years. Right?

是的。你知道，这些无晶圆厂芯片公司设计芯片的方式是：他们先完成设计，然后送到无晶圆厂公司生产原型，再送回测试。

Yeah. The way that, you know, in the with these fabless chip companies, the way they would design chips is they would work on the design. They would send them over to the fabless company. The fabless company would produce some prototypes. They'd send them back.

他们测试后，会来回反复几次。

They'd test them. They go back and forth a few times.

你是说代工厂会生产它们？比如台积电、三星或者格罗方德这样的？

You mean the foundry would produce them? Like the TSMC or the Samsung or the global foundries or

现在关键的是，英伟达目前没有使用台积电，因为他们用不了。台积电只与最顶尖的客户合作，而英伟达还不够格。所以他们用的是二流的代工厂。这个过程耗时很长。等到最后确认设计无误时，才会进行所谓的芯片流片。

Now importantly, NVIDIA is not using TSMC at this point because they can't they can't. TSMC only works with the best and NVIDIA is not the best. So they're using, like, second rate foundries. And that process takes a long time. And then at the end of it, when you're sure you got the the design right, then you do what's called a tape out of the chip.

顺便说一句，我超爱这个术语。

I love this term, by the way.

这术语其实源自过去制作芯片光刻掩模时真的要用胶带粘贴的年代。很酷对吧？不过现在就是指最终确定设计。

It harkens back to literally, like, when you used to tape, you know, masks to, like, do the photolithography on the chip back in the day. So cool. But it just means finalizing the design.

但你们实际上会先用一些原型机测试。比如代工厂会回传样品说：'感谢提供设计，这是做好的芯片，你们可以开始测试了'。

But you actually do run it on some prototypes first. Like, the the foundry sends back some, you know, hey. Thanks for the designs. Here's the chip. You know, run your tests on it.

要确保所有功能都符合预期。要知道，完成这样一个完整迭代过程需要两年时间。没错。

Make sure everything does what you think it does. And, you know, that process takes two years to get a a full sort of iteration on. Yep.

所以他们觉得'我们搞不定这个'。于是黄仁勋就说：'听着，我们要这么做。我听说现在有种新技术，有些新机器能模拟芯片。我们就用这个技术，在软件里完全模拟我们正在设计的显卡芯片。'

So they're like, we we can't do this. So, like, Jensen's here like, here's what we're gonna do. I've heard about there's this new technology, some new machines out there that enable emulation of chips. And in our case, we're gonna use it to emulate the graphics chip that we're we're designing all in software.

而且，你知道，它确实有效。虽然都是初创公司，但它们确实存在。

And it, you know, it works. They're startups, but they exist.

问题在于用软件模拟时，你知道，速度会非常慢。比如玩游戏时，你盯着电脑或显示器看，画面每秒刷新30到60次。如果是职业玩家，可能要求每秒120帧。而这个模拟器每三十秒才运行一帧。所以他们得用软件调试这东西，以节省时间——毕竟现在每三十秒才能看到一帧。

The problem is when you emulate it in software, you know, it's like it's really slow. So, you know, when you play a game and you're looking at your computer or monitor or whatever, it's refreshing 30 to 60 times a second. If you're a professional gamer, you probably have it going at like a 120 times a second, you know, frames per second. This emulator runs at one frame every thirty seconds. So they're gonna have to debug this thing in software to save this time going at one frame every thirty seconds.

这简直疯了。太残酷了。他们基本上在做这种权衡：如果想在九个月内交付产品，就没时间让硬件实际运行。所以只能牺牲测试效率——比如运行图形测试时，我们需要特定输出结果，就得安排人盯着屏幕，每三十秒看一帧新渲染的画面，再对照测试验证输出是否正确。

It's just insane. That's brutal. They're basically making this trade off of, okay, if we wanna ship something in nine months, we don't have time to actually have it execute on the hardware. So we're going to make the trade off of our testing being mind numbing, like running whatever our graphics tests are where we're looking for like this certain specified output. We need to plant someone in front of a screen to watch the new frame render once every thirty seconds and look against some tests to verify that the output is correct.

如果验证通过，这个人完成了这种枯燥工作，就这么一直观察、观察、观察，那我们就能直接投入生产，连实体原型都不用做。

And if it is, and this person does that mind numbing work and sits there just observing and observing and observing, then we will go right to manufacturing without ever producing a physical prototype and ship that.

而这正是他们的做法。他们花了一百万美元就为了搞这个模拟器，包括硬件和软件。

And that is exactly what they do. They have spent a million dollars just to get the emulator, you know, hardware and software to to do this.

我觉得他们虽然有些收入，但这笔钱仍占公司银行存款的三分之一。

Which I think they had generated some revenue, but it was still, like, a third of the cash that they had in the entire bank account.

公司资金只够撑六个月。他们在几个月内完成了任务，然后联系代工厂——不确定是用联电还是台湾其他厂，不是台积电。总之完成设计后就直接投产了。

They go down to six months until they're cash out in the company. They get it done in a few months and then they call up their foundry. I don't know if they're using, United or or one of the one of the other foundries in Taiwan, not TSMC. Like, alright. We tape this thing out, send it to production.

然后代工厂就说，你们确定要这样？他们说，是的，我们确定。生产10万片吧。

And the foundry's like, you guys sure about that? They're like, yep. We're sure. Make, you know, 100,000 units.

如果我没记错的话，英伟达基本上是那款仿真软件的唯一客户。那家初创公司还没完全证明自己，但英伟达表示，我们真的别无选择。

If I'm remembering right, I think NVIDIA basically was the only customer of that emulation software. Like, that was a startup that really wasn't fully proven yet, but NVIDIA was like, look, we literally have no options.

没错，他们是唯一客户。后来那家公司倒闭了，太疯狂了。所以他们设计的芯片...

Yeah. They were the only customer. And then that that company went out of business after. It's wild. Well, and so the chip they designed.

现在他们的优势在于——这种做法简直疯狂。显然他们别无选择。但优势在于，他们设计芯片时采用的技术假设与所有竞争对手相同。而竞争对手的设计方案需要18到24个月才能上市，英伟达却能在6个月内推出同代产品。

So now the advantage, like this is lunacy what they're doing. Obviously, they have to do it because their back is against the wall. The advantage of this though is they are now designing this chip with, you know, the same set of assumptions about what, you know, technology is available as all their competitors. But their competitors are working on those designs, they're not gonna be able get them out for like eighteen to twenty four months. NVIDIA is gonna get this same, you know, generation of design out in six months.

这款芯片叫Riva 128。它完全是个性能怪兽，各方面都堪称猛兽。

So this chip is called the Riva one twenty eight. That's what they call it. It is a freaking beast and it is like a beast in every sense of the word.

体积很大。

It's big.

确实很大。而且相比市面其他产品性能极其强悍。

It's big. It's extremely powerful relative to anything else on the market.

比任何客户所要求的都要强大得多。

Like more powerful than any customers are telling them they want.

是的，强大太多了。强大得太多太多太多太多了。不过，你知道，这也有缺点。能力越大，责任越大。

Yeah. Way more powerful. Way way way way way more powerful. But, you know, it comes with some downside. With great power comes, you know, responsibility.

因为他们构建这个东西的方式，它几乎没法用。有很多问题。我忘了具体数字，但当时Direct3D大概有24、25种不同的技术实现方式。

Because they built this thing in such a manner, it like barely works. Like there's a lot of stuff wrong with it. I forget the exact number of this, but, like, essentially, Direct three d at the time had something like, let's call it, like, 24, 25 different ways, like, different sort of techniques.

是混合模式吗？

These are the, blend modes?

对，我想就是混合模式。而Reva只能支持其中大约三分之二

Yep. I think that's what it was, blend modes. And the Reva only works with about two thirds

的模式。

of them.

剩下三分之一会直接崩溃，根本没法用。

Like, one third of them just, like, freaking crashes. Like, it just doesn't work.

我原本想得更糟。但基本上，我觉得英伟达不得不发起一场运动，四处游说所有不同的开发者，就像在说：'拜托，除了这八个功能，你们还需要什么？得了吧，你们真会用那些花哨的功能做什么？'

I thought even worse than that. But, basically, like, I I think NVIDIA had to launch a campaign going around to, like, all the different developers and being like, come on. What do you really need more than these eight for? Come on. What are you really gonna do where you need to use that fancy stuff?

帮我们个忙。这一代芯片就用这八个功能，效果棒极了。你会爱上它们的，它们太出色了。就用这些吧。

Do us a favor. For this generation of the chip, these eight work great. You're gonna love them. They're so good. And just use those.

好的。这简直太棒了，因为人们确实这么做了。他们从中了解到市场的需求。你看，英伟达最初的想法是：'我们要打造所有这些技术，我们要引领市场。'

Okay. So this is so so so great because people do it. And so what they learn from this, like, learn about the market. You know, the first iteration of NVIDIA, we're gonna build all this technology. We're gonna drive the market.

他们对市场一无所知，只是凭假设猜测人们想要什么。但现在他们真的走出去，黄仁勋亲自去说服开发者这么做，而开发者们都照做了。为什么？因为唯一重要的是性能。

They didn't know anything about the market. They were just making all these assumptions about what people wanted. But now they're actually going out and Jensen's going to these developers trying to convince them to do this and they all do it. Why do they do it? Because the only thing that matters is performance.

消费者会根据图形质量购买硬件和游戏。这就像第一次发现这个真理。人们愿意为了性能做出很多妥协。英伟达是第一个明白这点的，因为他们必须四处奔走说服开发者，而开发者们都接受了。

Consumers are gonna buy hardware and games based on the quality of the graphics. This is like being discovered for the first time. And so, like, people are willing to make a lot of compromises in, you know, service of performance. NVIDIA is like the first one that figured this this out because they have to go around and do this and developers all get on board.

明确地说，这是因为消费者在购买显卡时自己做决定，对吧？

And to be clear, it's because the consumer's making the buying decision, right, on what graphics card they buy?

这是一个完全相互关联的体系，消费者在做所有决定。需求就在那里——消费者决定购买什么硬件，这就是英伟达的生意本质。

It's it's a completely interrelated system where the consumer is making all of the decisions. That's where the demand is. The consumer is deciding what hardware to buy. That's what NVIDIA's business is.

无论是从OEM购买整机，还是单独购买显卡自行安装，他们都在决定电脑里装什么显卡。

Whether they're buying it as a fully, like, built computer from the OEM or whether they're buying the card to put in later themselves, they're making a decision on what graphics card goes in the computer.

没错。游戏开发者也在决定支持哪些显卡。他们要考虑目标消费群体——这款游戏会在哪些玩家的设备上运行？玩家是否需要至少达到x性能级别的设备才能流畅运行我的游戏？

Exactly. And the game developers are making decisions on what graphics cards to support Right. And how to build their games with, like, the assumption of what's my target market of consumers. Like, who do I think will this game run on? Do you need to have at least an x level performance rig in order to run my game or run my game in its fullest form?

所以开发者会预判游戏上市时市面上会有哪些显卡。他们会选择性价比最高的产品。基本上必须瞄准主流市场。如果我们测试后发现某款显卡在每元性能或每瓦性能上最优——也就是效率最高的显卡——那么玩家就会选择它。

So the developers are premeditating what graphics cards are going to be out in the market when their games launch. And they're saying Yes. It's gonna be the most performant one at the right price point. So whatever the mass market is, we kinda have to target that. And if you're telling us and we're gonna test it and it turns out that yours is the best performance per price or performance per watt or whatever, it's the most efficient card, then people are going to buy that one.

因此我们必须瞄准它。

And so we must target it.

那张显卡的玩家会购买我的游戏。我记得几年前有过这种情况。有款游戏叫《孤岛危机》（Crysis），记得吗？

That card and they're gonna buy my game. I mean, I remember, like, this is a few years later. This is a, you know, a trope that happened. There was a game called Crisis, c r y s I s. Remember this?

哦，记得。《孤岛危机》和《孤岛惊魂》有什么关系？

Oh, yeah. What's the relationship between Crisis and Far Cry?

不，不是。《孤岛惊魂》是第一部作品。后来用Crysis引擎开发的《孤岛危机》剧情超级复杂。我记得《孤岛惊魂》刚推出时——大概是2005年左右——画面效果简直难以置信。

It was oh, no. Far Cry was the first game. Yeah. The Crysis engine and then Crysis also it was super convoluted. Basically, my perception of this thing was when this came out, when Far Cry came out, this was like mid two thousands, the graphics were unbelievable.

难以置信。如果你有一台足够强大的设备来运行它，简直不可思议。游戏本身完全是个垃圾。我估计自己玩的时间从没超过十分钟。

Unbelievable. And if you had a rig powerful enough to run it, like, just unbelievable. The game itself was total crap. Like, I don't think I ever played more than ten minutes of it.

我很确定如果你的电脑不支持，网上会有很多玩家录制的视频，比如堆叠上千个汽油桶然后射击。由于显卡性能不足，电脑直接卡死。这就是《孤岛惊魂》在低配芯片上的崩溃模式。

I'm pretty sure if your computer didn't support it, there was all these videos that people would record of, like, building a tower of, like, a thousand gasoline barrels and then shooting it. And because it was too complex for their graphics card to handle, their computer would just freeze. That was the failure mode of Far Cry with non performing chips. This is how

硬核游戏产业就是这样进化的。《孤岛惊魂》卖出了海量软件和硬件，就因为人们想体验那种级别的画质。开发商也开始明白这个道理，他们决定：只要能实现效果，哪怕只支持八种混合模式也在所不惜——图形性能才是最重要的。

the hardcore gaming industry evolves. Like, Far Cry sold so much software and so much hardware just because people wanted to experience that to attempt to experience that level of graphics. And so that's what the developers are starting to figure out. They're like, alright. Well, if you can ship this thing, we'll use only those, you know, eight blend modes or whatever, like whatever it takes because we want, you know, graphical performance is the most important thing.

这招很管用。Riva 128显卡四个月内卖出一百万台。哇，我该查查它的建议零售价的，但这营收确实惊人。

So it works. They sell 1,000,000 units of the Riva one twenty eight within four months. Wow. I should have looked what the MSRP was of it, but that is a lot of revenue.

是啊，不开玩笑。这是哪一年的事？

Yeah. No kidding. What year was this?

1997年。

This was 1997.

明白了。那是个有趣的时代，互联网已经兴起，距离.com泡沫破裂还有几年。PlayStation一代已经上市，但PS2应该还没发布。

Okay. So we're it's an interesting era. Like, the Internet is a thing. We still have a few more years till the .com bubble crashes. PlayStation one is out, but PS two is not out yet, I think.

是的，PlayStation一代。自此，游戏市场大致分化为两个方向：一方面是标准化的主机市场，你知道，一切都将按预期运行；另一方面则是硬核PC游戏市场，虽然用户基数较小，但收入潜力巨大，因为玩家愿意在这方面投入大量资金。最终，NVIDIA洞察了PC游戏市场的这些动态，并在公司内部建立了一套流程，能够在六个月内完成每一代新硬件的设计与交付，而整个行业的周期通常是18到24个月。

Yep. PlayStation one. And with that, the gaming market kinda bifurcated into like sort of the, you know, the console market which was standardized and you knew it was all gonna work. And then the the hardcore PC gaming market which just had so much revenue potential even though it was smaller in terms of numbers because people are willing to spend so much money on this stuff. So at the end of this, NVIDIA has now figured out these dynamics of the PC gaming market and they now have a process within the company to design and ship each next generation of their hardware in a six month timeline, while the rest of the industry is on an eighteen to twenty four month timeline.

需求是发明之母。

Necessity is the mother of invention.

说这是巨大的突破简直是本世纪最轻描淡写的说法，巨大无比。对这个市场而言意义重大，但当时根本没人意识到这一点。连黄仁勋都没预见到，所有人都没料到。他们现在基本上每代硬件性能都翻倍，并且每六个月就推出一代新品。

To say this is huge is like understatement of the century, huge. And it's huge for this market but nobody even saw this at the time. Like Jensen didn't see this, nobody saw this. They're now shipping relatively, you know, doubling essentially the performance in each generation with their hardware. And they're shipping it every six months.

想想摩尔定律对吧？摩尔定律原本说的是芯片上晶体管数量——即市场某价位能获得的算力——每18到24个月翻一番。而NVIDIA从1997、98年开始的周期里，他们每六个月就能让市场某价位获得的性能翻倍。

And you think about Moore's law, right? Like Moore's law was that the number of transistors on a chip equating to the compute power available at a given price point to the market would double every eighteen to twenty four months. NVIDIA is now on a cycle starting in 1997, 1998, where they are doubling the performance that they are delivering at a given price point to the market every six months.

这太迷人了。而且他们与CPU制造商竞争的维度完全不同——说来神奇，我们这期节目都聊了一小时才提到这点。GPU的魔力在于其高度并行化。对于上过低阶计算机课程的人，CPU的工作原理大概是：时钟每跳动一次，执行一条指令，操作在CPU内部漫长的处理链中推进，本质上是串行处理。

It's fascinating. And And they're also competing on a different vector than the CPU manufacturers because and it's kind of amazing we've made it an hour into the episode and haven't talked about this yet. But the magic of GPUs is that they're very, very parallel. Like CPUs, for anyone who's taken a low level computing class, you sort of know that, like, every time the clock ticks, an instruction can sort of run, and things move through the sort of long chain of operations that can happen within the CPU. And it's advancing things serially through the processor.

这是串行处理。

It's serial processing.

它

It

可以从寄存器读取数据，或者将两个东西相加，但这一切都是串行处理的。

can read from a register or it can add two things together, but, like, it's all happening serially.

就像《我爱露西》里那个著名的场景，巧克力从工厂流水线上下来，CPU就像一个个包装工人，逐个包装每块巧克力。

It's like the, the I Love Lucy, you know, famous one where, like, the chocolates are coming down the factory pipeline, and you had the CPUS to, like, wrap each individual chocolate one and then the next one.

没错。图形处理的魔力就在于它超级适合并行处理。屏幕上需要输出的许多内容彼此独立，因此可以同时进行。

Yes. Exactly. And with graphics processing, like, the magic of it is that it's super parallelizable. Like, there's all these things that need to get outputted to the screen that do not depend on each other. And so you can do them independently.

所以他们竞争的维度其实是——虽然要多年后才能实现——不断增加核心数量，或者寻找更多方法同时执行更多指令来实现任务并行。当时人们认为并行化的主要用途就是图形处理。这个观点我们先放一边，但值得了解的是，他们正在探索如何更快地并行处理更多任务。

And so the vector that they're competing on is really like, oh, we can and that it would be years before they would really get to this, but add more and more cores or find more ways to execute more instructions simultaneously to parallelize these tasks. And I think at the time, people thought, really, the only big use case for parallelization is graphics. Let's put a pin in that for now, but it's worth knowing the thing that they're doing is figuring out how to process more things in parallel faster.

是的。当时英伟达制造的显卡特别擅长并行点亮屏幕像素，每秒30、60甚至120次，根据来自CPU的游戏或图形程序提供的图像数据。游戏开发者使用微软Direct3D（后来变成DirectX）或开源竞争对手OpenGL进行开发。所有逻辑运算都在CPU领域完成，这就是为什么那个时代的游戏——想想PS1、PS2、N64主机游戏或同期PC游戏——

Yes. So graphics cards like NVIDIA is making at this point in time are really good at in parallel lighting the pixels on a screen, you know, thirty, sixty, a 120 times a second with the images that are being fed to them from, like, the game or the graphics program, which is living all in the CPU land. So, like, you're a game developer, you develop in, you know, Microsoft Direct three d becomes DirectX or OpenGL is the open source competitor to this. You know, all that logic is really happening in the CPU realm and what that means is like if you think back to games from this time, you know, think console games, PlayStation one, even PlayStation two, n 64. When you look at the graphics in those games or PC games from

它们的画面看起来都差不多。

the time too, they're all kind of the same.

完全一样对吧？所有光照效果都是预先设置好的。游戏开发者搭建场景时，你绝不会看到角色举着火把奔跑时，火把光线实时影响周围环境的效果。一切都是预先设定好的。

They're all the same. Right? All the lighting, like, the lighting it's all like predone. So like when you're a game developer, you set the scene, you'd never see like a character running around carrying a torch and that torch light like impacting the rest of the environment. It's all set in advance.

比如，GPU层面在屏幕显示上没有任何智能处理，它只是点亮像素而已。

Like, no intelligence is happening in the GPU level with the screen. It's just lighting up the pixels.

基本上，为了让开发者更轻松，软件开发工具包被设计得如此高层级，以至于你实际上无法获得足够的控制权来使游戏风格独特。你只能把物品摆放在屏幕上。

Basically, in order to make it easy for developers, the software development kit is written at such a high level that you don't really get enough control to make your game stylistically different. You just get to lay out the items on screen.

全都一个样，都是平面的。也许你可以编程实现，比如硬编码，比如根据一天中的时间变化来改变视觉效果。但你是通过硬编码决定它们的样子，没有实时计算发生。

It's all the same. It's all flat. Maybe you can program that, like, hard code that, like, oh, time of day might change and, like, that might change the way things look. But you're hard coding, like, what they look like. No computation is happening.

没错。如果你今天玩个游戏，哪怕是最基础的手游之类的，你到处都能看到动态光影效果——这个我们稍后会详细讨论。所以这就像...你知道的，GPU虽然是非常重要的商品，但它们终究只是商品。这里面并没有太多智能成分。是的。

Right. If you're playing a game today, even the most basic, you know, mobile game or whatever, you're seeing dynamic lighting and shading, which we'll get into in a sec, all over the place. So this is still like in the you know, GPUs are like a really really important sort of commodity, but they're a commodity. There's not a lot of smarts happening here. Yep.

不需要编程。但英伟达已经解决了这个问题。他们现在能以六个月为周期推出新品，开始真正抢占巨大市场份额。现在很多人开始以积极态度关注他们了。

No programming. But NVIDIA has figured this out. They can now ship on a six month time cycle. They're starting to like really take huge market share. Now, a lot of people start paying attention to them in a good way.

台积电当年连黄仁勋的电话都不接。有个非常精彩的故事——你看过台积电三十周年纪念活动吗？

TSMC that wouldn't even return Jensen's calls back in the day. There's this amazing amazing story. Did you watch the TSMC thirtieth anniversary

我看过。

I did.

庆祝？这就像在YouTube上看了三个小时。

Celebration? This is so it's like three hours on YouTube.

这值得稍微提一下。台积电的张忠谋影响力有多大。他把英伟达Arm的CEO们都请上了台。Arm、ASML、高通和博通。

This is worth a brief aside. This is how much pull Morris Chang from TSMC has. He gets the CEOs on stage of NVIDIA Arm. Arm, ASML, Qualcomm, and Broadcom.

是啊。AMD的苏姿丰好像没在场。

Yep. I don't think Lisa from AMD was there.

没错。基本上除了AMD之外，台积电生态系统的支柱企业都来了。张忠谋亲自担任采访者，看他主持非常有趣。

No. It's basically everyone but AMD of the sort of pillars of the TSMC ecosystem. I mean, Morris is playing interviewer. Like, it's very entertaining to watch him.

这就像是对张忠谋和台积电的庆典。太棒了，真的太棒了。是的，在与黄仁勋的环节中，他们讲述了英伟达如何成为台积电最大客户的故事。

It's like a celebration of Morris and of of TSMC. It's amazing. It's amazing. Yes. So in the section with Jensen, they tell the story of how Nvidia, at this point it's gotta be TSMC's biggest customer.

他们就像永远绑在一起似的。Riva 128大获成功后，黄仁勋写了封实体信，寄给台湾的张忠谋。

I mean, they've been like tied at the hip forever of how this all came to be. After the Riva one twenty eight hits and it's become a big success, Jensen writes a letter to more like a physical letter, addresses it to Morris Chang in Taiwan.

因为他通过销售人员根本联系不上。

Because he can't get in touch through any of the like salespeople.

没错。完全正确。他们之前都无视了他，这也是应该的，因为那是一家被遗弃的初创公司，淹没在创业海洋中。这封信辗转到了张忠谋手中，他在台湾拆开阅读后，做出了最符合张忠谋风格的决定。

Exactly. Exactly. They've all just been ignoring him as well they should because they were a, you know, left for dead startup in a sea of startups. The letter gets to Morris, he opens it, he reads it in Taiwan. He does the most Morris Chang thing possible.

他当场就给黄仁勋打了电话。据英伟达办公室流传的故事，电话铃声响起时，他们正疯狂赶工准备出货这批刚下线的Riva 128显卡。由于全是未经测试的新品，员工们正在办公室手工检测每块显卡。

He calls up Jensen on the phone right there. And the phone rings as they tell the story in the NVIDIA office. This is in the middle of their trying, like, mad scramble as a startup to ship these Riva one twenty eights that are coming in. They're testing them all by hand in the office because none of this stuff was it's fresh off the line. It's not been tested.

场面一片混乱。黄仁勋接起电话问：'哪位？'张忠谋说：'你好，我是台积电的张忠谋。我收到了你的信。'

It's chaos. Jensen picked up the phone and like, yeah, who's this? Morris is like, hello. This is Morris Chang at TSMC. I got your letter.

张忠谋回忆说电话那头沉默了几秒，随后听到黄仁勋大喊：'都安静！张忠谋来电话了！'简直不可思议。

And Morris says that there's like a silence on the other end for a couple seconds and then he hears Jensen yelling, everybody shut up. Morris Chang is on the phone. Amazing.

就这样，台积电成为了英伟达的芯片制造商。

And that's how TSMC became the manufacturer of NVIDIA chips.

没错。次年两家公司签署了多年期大单，台积电正式成为英伟达主要代工厂，合作延续至今。黄仁勋与张忠谋私交甚笃，这份协议对双方都具有里程碑意义。有了顶级代工伙伴和独特的芯片研发流程，英伟达开始突飞猛进。

Yep. The next year, the two companies sign a huge multi year deal for TSMC to become the primary foundry for Nvidia and still are today. Jensen and Morris are super close. It's a landmark landmark deal for both companies. So with now an actually really good foundry as their partner and this super unique chip development process, NVIDIA just keeps accelerating.

1999年他们进行了产品线更名——最初使用NV1，后来是Riva 128。公司内部举办了征名活动，获胜方案是'Geometry Force（几何原力）'，最终缩写为GeForce。如今但凡懂显卡的人都知道，英伟达GeForce系列至今仍是游戏显卡旗舰品牌，堪称游戏硬件领域最受尊敬的品牌之一。这一切都源于199年推出的首款GeForce显卡——GeForce 256。

So in 1999, they rebrand their products. You know, they'd use the NV one first and then this Reva one twenty eight. They actually run a little contest of what they should name the products and the winning name is Geometry Force. Force is with you, which they shortened to GeForce, which anybody who knows who, you know, buys graphics cards, the NVIDIA GeForce, still the brand name they use for their gaming cards today and is probably the most one of the most respected, you know, brands in the gaming ecosystem. And it's because this card that they ship, the first GeForce in 1999, it's the GeForce two fifty six.

它的性能太强大了，图形处理能力比市面上其他产品高出五倍。

It's so powerful. It has five x better graphics performance than like anything else on the market.

他们称这是第一款GPU对吧？不是说他们发明了GPU这个概念吗？

And they call this like the first GPU. Right? Don't they say like we're inventing the GPU?

他们称之为GPU。在此之前GPU这个术语并不存在，只有图形卡、图形芯片这类说法。

They call it a GPU. Before this, the term GPU didn't exist. It was these were graphics cards, graphics chips.

我记得索尼在PlayStation上好像用过类似说法，但没人把这个概念作为营销重点。

I think Sony had, like, sort of used it about the PlayStation, but no one's marketing this idea.

所以他们将其营销为图形处理器单元。一方面这像是营销噱头，另一方面这个声明其实暗藏玄机。为什么这么说？黄仁勋和英伟达到底想表达什么？

So they market this as the graphical processing unit. Now on the one hand, that's like sort of like marketing bravado. On the other hand, that is like a very loaded statement to make. And why so? What does Jensen and NVIDIA mean by this?

说到英特尔，提到芯片人们就想到英特尔对吧？提到半导体就想到英特尔。当时英特尔的整体战略基本就像今天的生物科技公司，好比大型制药企业。或者换个说法，就是微软'拥抱、扩展、消灭'策略的另一个版本。

So Intel, you know, you think chips, you think Intel. Right? You think silicon, you think Intel. Intel's whole strategy at this point in time was basically they're almost like a biotech companies today, like one of the big pharma companies. And, or or put another way, was another version of the Microsoft embrace, extend, extinguish thing.

他们会观察所有外围设备——声卡、网卡、显卡，我们讨论过的所有这些产品。他们会放任这些产品百花齐放，表现得很大度：'随便用我们的主板PCI插槽，没关系，我们是开放生态系统。'

They would see there are all these peripherals, sound cards, networking cards, all the graphics cards, all the stuff we've talked about. They would let all these flowers bloom be like, yeah yeah yeah. Just plug into the PCI slots on our motherboards. No big deal. We're open ecosystem.

我们希望每个人都能蓬勃发展。然后他们会观察哪些外围设备获得了消费者的青睐，接着就把它们变成主板上的一个组件。

We want everybody to flourish. And then they would see which of these, you know, peripherals got consumer traction, and then they would just turn them into, you know, a component in the motherboard.

于是开启了这样一个时代：人们可以购买带有英特尔主板和集成显卡的PC。

And thus began the wave of being able to buy a PC with, an Intel motherboard and integrated graphics.

在那之前还有集成声卡、集成网卡。记得吗？做这项研究真是太有趣了。还记得创新公司和声霸卡吗？哦，是的，我记得买过一大堆那玩意儿。

Well, and before that, you know, integrated sound, integrated networking. Like, remember, oh, was so fun doing this research. Remember the company Creative and the Sound Blaster cards? Oh, yeah. I remember buying tons of that stuff.

到某个阶段后，你就不再购买声霸卡了，对吧？

Like then at a certain point, you stopped buying Sound Blaster cards. Right?

你会觉得，主板已经满足了我90%的需求，为什么还要额外花钱买独立设备呢？

You're like, oh, the motherboard does 90% of what I need it to do, and why would I spend extra money on a separate thing?

没错。英特尔就坐观其变，看着这一切发生，然后将其整合。对初创公司来说游戏就结束了。

Exactly. And so Intel, they just sit back, they'd watch all this happening, They'd integrate it. Game over for the startups.

当然也存在需要专业设备的情况。比如我记得买过特殊网卡，因为我的Mac 8500（好像是这个型号）主板集成的网卡速度不如专门的PCI网卡快。显卡也类似，对大多数人来说集成显卡就够用了，除非你是游戏玩家——这时你会自己购买显卡，或者直接从OEM厂商那里定制整机时选配。

And there was, like, reasons for specialized stuff. Like, I remember buying a special network card because the integrated networking capability of the motherboard on my I don't know what it was, a Mac 8,500 or something wasn't as fast as, like, if you bought a dedicated PCI card that could be a faster networking card. And graphics cards would sort of become that same thing where the integrated graphics for most people was good enough unless you were a gamer, in which case you'd go buy your own graphics card or you'd buy it directly from the OEM when they were making the computer and shipping it to you.

但等上一两代人。即使你对家庭网络性能要求再高，也不会买那种独立网卡了，简直该被淘汰。

But wait a generation or two. Even if you have the most demanding performance for home networking, you're not buying that separate networking card, like get out of here.

这些业务就像是死胡同。

These things are like dead end businesses.

显卡没理由不走上同样的道路。所以黄仁勋和英特尔跳出来说，我们是图形处理单元，是GPU。这简直是对英特尔和整个CPU主导世界的巨大嘲讽。

And there's no reason why graphics cards wouldn't be the same. So Jensen and Intel coming out and being like, we're a graphical processing unit. We're a GPU. It's a big middle finger to Intel and this whole CPU dominant world.

当时这其实还不成立。它并不像CPU那样是真正的处理单元——人们能为它编写软件，从而为用户创造截然不同的体验。没错，但是...

And it really wasn't true yet. It wasn't a processing unit in the same way that a CPU was a processing unit where it was people could write software for it in a way that created a meaningfully different experience for people using the software. Yep. But

这就是黄仁勋展现战略大师本色、英伟达大放异彩的时刻。接下来几年里，他们精心布局的多项举措接连爆发。首先英伟达上市，当时Rebo 128大获成功，新款GeForce 256也供不应求。

this is where, you know, Jensen is just such a master strategist and Nvidia was so great. Like, this whole kind of orchestration of a bunch of things all hit over the next couple years. So first, Nvidia goes public. You know, they've now shipped the Rebo one twenty eight was a huge hit. This new GeForce two fifty six flying off the shelves.

他们在1999年初以6亿美元市值上市。相比红杉和Sutter Hill那轮融资后600万美元的估值，实现了100倍回报。这为公司带来了更多资金。与此同时，他们正与微软秘密洽谈——当时微软正在开发Xbox项目。

They go public in beginning in 1999 at a $600,000,000 market cap. So a 100 x return from the $6,000,000 post money valuation on the Sequoia and Sutter Hill round. That gets them, you know, some more capital. And then behind the scenes, they're working, they're in talks with Microsoft. Microsoft's got a secret project that they're working on at this time, the Xbox.

嗯，我们在索尼专题和节目里多次讨论过。微软找到英伟达说：希望你们成为Xbox图形处理器GPU的核心供应商。双方达成了每年5亿美元的巨额协议，其中包含2亿美元的预付款。

Mhmm. Which we talked about a lot on the Sony episode and so many times on the show. And Microsoft comes to Nvidia and like we want you to be a key supplier of the graphics at the GPU for the Xbox. And they do a huge, huge deal. $500,000,000 a year deal for NVIDIA to supply the graphics for the Xbox with a $200,000,000 advance.

他们使用的芯片是英伟达正在研发的这种革命性新芯片的改良版本。这不像史蒂夫·乔布斯那种风格。黄仁勋谈论这个时的语气倒像乔布斯。GeForce 3首次在GPU上实现了可编程着色器和光照效果。我们刚才讨论的所有功能——比如GPU大规模并行处理可以照亮所有这些像素——但本质上它只是执行预先硬编码的固定光照指令。

And the chip that they use is a modified version of this incredible new chip that NVIDIA is working on. It's not like Steve Jobs here. Jensen sounds like Steve Jobs talking about this. The GeForce three, which introduces for the first time programmable shaders and lighting on the GPU. Everything we just talked about though, like, the GPU massively parallel can light all these pixels, but it's essentially just taking instructions that are pre, you know, hard coded baked in on what the lighting's gonna look like.

现在你可以为这些GPU编程，在游戏和3D图形中实现动态计算的光照效果。

Now you can program for these GPUs and you can make dynamic lighting in games and three d graphics that is calculated.

这具有颠覆性意义。可以这样理解：那些所谓的'GPU'实际上是固定功能的图形加速器。它们能将纹理映射到多边形上，但无法实现你所说的功能，大卫——比如自定义光照效果，以及在GPU层面编程控制各种效果。这当然很酷，因为它开启了全新的用户体验浪潮，每个游戏开发者都能在游戏中留下自己的风格印记。但这彻底改变了计算机架构的范式，突然间GPU变得可编程了。

This is game changing. The way to think about it is those GPUs in quotes were fixed function graphics accelerators. So they would be able to map textures onto a set of polygons, but you couldn't do the thing that you're talking about, David, custom lighting, a lot of that sort of stuff to to actually program at the GPU level what is happening. And so this is like, of course, it's cool because it's a wave of new consumer experiences that can happen because every game developer can kind of stylistically put their own stamp on games. But it's a totally different metaphor for the computer architecture, where suddenly you can program a GPU.

我想这就是他们称之为GPU的原因，这与传统显卡完全不同。

And I guess that's why they're calling it a GPU, and this is different than a graphics card.

英伟达为此同步开发了CG语言。他们实际上扩展了C语言，增加了图形库和功能，可以直接为GPU编程图形、光照和着色器。这让那些营销口号——比如'GeForce 256是真正的GPU'——终于名副其实。

And NVIDIA develops in conjunction with this. They call it CG. Literally, like, they extend the c programming language with graphics, libraries, and capabilities to directly program graphics and lighting and shaders for the GPU. So this makes, you know, that sort of like marketing, you know, oh, this GeForce two fifty six, it's a GPU. Now it's real.

这是一种智能的图形处理单元，其重要性虽然暂时还比不上CPU，但它就像插在地上的界碑，宣告这绝非声卡之流，不会被商品化浪潮淹没。

Like, this is a graphical processing unit that is intelligent, that is every bit as you know, maybe not every bit as important as the CPU yet, but, like, this is, like, the stake in the ground of, like, this is no sound card. This is not gonna get commoditized.

你知道这是GeForce FX吗？还是说GeForce FX是面向PC的类似版本？这个问题问得好。

Do you know if this was the GeForce FX or if the GeForce FX was a similar version of this that was available to PC? That's a good question.

这是GeForce 3，也就是这个的PC版本。

It was the GeForce three was the the PC version of this.

好的。转向可编程着色器是场押上公司命运的豪赌，这是黄仁勋对'如何摆脱低利润业务、做出独特创新'的解答。我几乎可以肯定，他们当时距离再次耗尽资金只有几个月之遥，因为这个全新产品线需要极其激进的团队扩张。

Okay. This move to programmable shaders was a bet the company move, and it was Jensen's answer to how do we get out of this commodity business and do something unique and different. And I'm pretty sure they were, like, months away from cash out again by pulling this move because of how aggressively they had to staff this, like, very new type of product they were inventing.

没错。这其实回归了公司最初那种堂吉诃德式的愿景——我们要创造一个新产业，开发配套的API和SDK，包办所有环节。如今他们真的做到了，而且这次是与微软合作而非对抗，确实是步好棋。

Yeah. I mean, this is the you know, back to that original sort of Quixotic vision for the company of we're gonna create an industry, we're gonna create the APIs, the SDK to interface with it, we're gonna do all this. Like, now they're doing it and they're doing it with Microsoft this time instead of, like, against Microsoft. So, like, a plus move there. Yeah.

但确实，这个项目的资本投入规模太惊人了。此时英特尔才意识到：我们可能遇到麻烦了。

But yeah, like the amount of capital investment that went into this was enormous. So at this point, Intel's like, we might have a problem here.

没错。要把这些人做的东西直接集成到我们的主板上，难度远超预期。

Right. It's gonna be more difficult than we thought to just take whatever these people are doing and integrate it directly into our our motherboards.

是的。最具讽刺意味的是，黄仁勋更进一步——他与AMD达成了重大合作。

Yep. And, irony of ironies, Jensen presses this even further. He does a big partnership with AMD.

需要说明的是，你提到AMD时，要知道虽然现在AMD和英伟达是GPU领域的主要竞争对手，但当时还不是。那时的AMD主要生产CPU处理器，与英特尔竞争。

It's worth knowing here when you're saying AMD because people probably know AMD and NVIDIA are big competitors today in the GPU world. Not yet. Right. AMD primarily made CPUs at this point. They made processors and competed with Intel.

他们当时还没有收购ATI，也就是现在Radeon业务的来源。这就是他们如今所有的图形业务。

They hadn't yet bought ATI, which is where the Radeon business comes from. That's all the graphics stuff that they do today.

没错。那时候ATI是英伟达的第二大竞争对手。实际上也是个了不起的故事。它是一家加拿大公司，八十年代创立后转型做显卡，完全不同的路径。我觉得这里面有值得学习的经验，可以放进创业手册里讨论。当所有风投都在资助那些90年代的硅谷初创公司做3D显卡时，最终存活下来的只有英伟达——它经历了地狱般的历程。

Yeah. ATI at this point was the number two competitor to NVIDIA. Actually, an amazing story too. It was a Canadian company, started in the eighties and pivoted into graphics cards, like very different, you know. I feel like there's a lesson in here, right, we could talk about this in playbook, but like, when all the VCs funded these 90, you know, Silicon Valley startups to go make graphics cards, three d graphics cards, the only two surviving ones were NVIDIA, which went through this hellish journey.

还有这些完全游离在生态系统之外的加拿大人，他们以更自力更生的方式进入这个领域并逐渐壮大。

And then these Canadian guys that were, like, totally out of the ecosystem and, like, did it sort of more in a boot more bootstrapped way and evolved into the space.

黄仁勋对此有句名言，他后来在斯坦福大学演讲时说：'当技术发展如此迅猛时，如果你不自我革新，就是在缓慢死亡。不幸的是，这个死亡速度遵循摩尔定律——这是我们已知最快的速度。'

Jensen has a great quote about this, and he's giving this lecture at at Stanford years later. And he says, when technology moves this fast, if you're not reinventing yourself, you're just slowly dying. You're slowly dying. Unfortunately, at the rate of Moore's Law, which is the fastest of any rate that we know. Yep.

这句话清晰揭示了他对英伟达必须进行三次彻底转型的思考，必须全盘押注、承担风险。否则就会成为那89家失败公司之一。

It's so clarifying of how he thinks about why NVIDIA needed to do these, like, three complete transformations of the company, bet at all, risk at all. Because if you're not, you're one of those 89 companies.

正是。所以英特尔当时大概在想：'见鬼，我们可能遇到麻烦了。不过这对英特尔来说不算真正的危机。'

Exactly. So Intel's like, holy crap. We might have a problem on here. Not not a pro like, this is not a problem for Intel.

这只是他们必须应对的问题，而非其既定战略的组成部分。

It just is a a thing they're gonna have to deal with instead of it being part of their extinguished strategy.

没错。英特尔现在已经习惯了，就像微软那样。哦，当然。你想开发WordPerfect？我们会让你这么做的。

Right. Intel is used to at this point just, you know, like Microsoft at this point. Oh, sure. You know, you wanna go make WordPerfect? We'll we'll let you do that.

我们会看到这些优秀的应用，然后自己开发一个。英特尔就是这么做的。而现在这是第一个例子，说明英特尔独自做这件事会遇到麻烦。所以他们最初推出了自己的独立英特尔显卡，你知道吗，GPU，作为独立显卡竞争。哇哦。英特尔以前可没这么干过。

We'll see these great applications and then we'll go make our own. That's what Intel's doing. And now this is the first example of, like, Intel's gonna have some trouble doing this on their own. So they actually, at first, come out with their own dedicated Intel graphics, you know, GPUs, graphics cards competing as separate cards Woah. Other than Intel had ever done that.

我的意思是，也许我说得不太恰当，但据我所知，这不是英特尔的常见策略。他们通常是集成——对，集成到主板和CPU里。而就在这个时候，大约1999年，他们推出了自己的外置显卡直接竞争。结果呢，这些显卡烂透了。

I mean, and maybe speaking out of turn here, but, as far as I know, I don't this is not a common strategy for Intel. It's usually integrate Yeah. Into motherboard and the CPU. They come out with their own external cards right around this time, like 1999, to directly compete. And, like, they suck.

这些可以说是历史上评价最差的显卡之一。

Like, these are, like, some of the worst reviewed graphics cards in history.

这完全不是你的核心竞争力啊。

Talk about not your core competency.