还有什么能阻挡英伟达？| 经济学人商业
1
思维导图：
Summer，女，QE在职，梦想能仗走天涯翻译/音乐/健康
2024年第一期翻译打卡营
中鸟票最后两天
中鸟票最后两天
中鸟票最后两天
4位一笔，3位二笔
30天录播课程
既有中译英，也有英译中。
从翻译的流程策略，
英汉语言特点，方法，类型，
背景知识到细致的语言点，
我们都逐一讲解以及答疑，让同学们吃透翻译。
点击下图，即可了解翻译打卡营详情！
2
Business | Schumpeter
经济学人商业 | 熊彼特
英文部分选自经济学人20240323期社论版块
Business | Schumpeter
经济学人商业 | 熊彼特
Can anything stop Nvidia’s Jensen Huang?
英伟达黄仁勋：无人能挡
He has become the generative-AI showman of our time
老黄或已成为当代生成式人工智能看板娘
Jensen Huang is a man on a mission—but not so much that he does not have time to tell a good story at his own expense. Last spring, when his semiconductor company, Nvidia, was well on its way to becoming a darling of generative artificial intelligence (ai), he and his wife bought a new home in the Bay Area. Mr Huang was so busy he could not spare much time to visit it before the purchase was completed. Pity, he admitted later, sneezing heavily. It was surrounded by plants that gave him hay fever.
黄仁勋肩负重任——但也还没重到连亲自讲个故事的时间都抹不开。去年春天，老黄的半导体公司英伟达发展正猛，大有成为生成式人工智能的宠儿之势，他和妻子在旧金山湾区新置了一套房产。他忙得脱不开身，房子交付前都没能挤出时间去实地查看。后来，当事人狂打喷嚏表示“现在就是后悔，非常后悔”。屋子周围的植物成功让老黄花粉过敏。
Mr Huang uses such self-deprecating humour often. When he took to the stage on March 18th for Nvidia’s annual developers’ conference, to be greeted by cheers, camera flashes and rock-star adulation from the 11,000 folk packed into a San Jose ice-hockey stadium, he jokingly reminded them it wasn’t a concert. Instead, he promised them a heady mix of science, algorithms, computer architecture and mathematics. Someone whooped.
黄仁勋很喜欢这种自嘲式的幽默。3月18日，黄仁勋走上英伟达年度开发者大会的讲台，迎接他的是震耳的欢呼声、闪烁的镁光灯以及摇滚明星一般的崇拜。他打趣地提醒挤满圣何塞冰球馆的11000多名观众，这不是一场音乐会，而是一场融合了科学、算法、计算机架构和数学的饕餮盛宴。一位观众激动地大喊起来。
In advance, Nvidia’s fans on Wall Street had dubbed it the “ai Woodstock”. It wasn’t that. The attendees were mostly middle-aged men wearing lanyards and loafers, not beads and tie-dyes. Yet as a headliner, there was a bit of Jimi Hendrix about Jensen Huang. Wearing his trademark leather jacket, he put on an exhilarating performance. He was a virtuoso at making complex stuff sound easy. In front of the media, he improvised with showmanship. And for all the polished charm, there was something intoxicating about his change-the-world ambition. If anyone is pushing “gen ai” to the limits, with no misgivings, Mr Huang is. This raises a question: what constraints, if any, does he face?
先前，华尔街上的英伟达粉丝将这场大会称为“人工智能界的伍德斯托克音乐节”。事实并非如此。毕竟与会者多为中年男子，颈戴挂绳脚穿乐福鞋，未佩戴任何首饰，穿的也不是扎染衣物。主角老黄倒还有些吉米·亨德里克斯（Jimi Hendrix）的风格：他身穿一件标志性的皮夹克，带来了一场振奋人心的表演。他擅长把复杂之事说得通俗易懂。他深谙表演艺术，在媒体面前即兴演出。除了十足的魅力，他还有着改变世界的雄心，令人着迷。如果真有人能毫无顾虑地探究生成式人工智能的极限，那便是黄仁勋了。问题来了：还有什么能阻挡老黄？
注释：伴随着大会影响力的与日俱增，该活动也被称为“AI界的伍德斯托克AI Woodstock”（指伍德斯托克音乐节Woodstock Rock Festival，在美国纽约州北部城镇伍德斯托克附近举行，是世界上最著名的系列性摇滚音乐节之一，最早举行于1969年）。
source: https://new.qq.com/rain/a/20240319A013RI00
The aim of the conference was to offer a simple answer: none. This is the start of a new industrial revolution and, according to Mr Huang, Nvidia is first in line to build the “ai factories” of the future. Demand for Nvidia’s graphics-processing units (GPUs), ai-modellers’ favourite type of processor, is so insatiable that they are in short supply. No matter. Nvidia announced the launch later this year of a new generation of superchips, named Blackwell, that are many times more powerful than its existing GPUs, promising bigger and cleverer ais. Thanks to ai, spending on global data centres was $250bn last year, Mr Huang says, and is growing at 20% a year. His company intends to capture much of that growth. To make it more difficult for rivals to catch up, Nvidia is pricing Blackwell GPUs at $30,000-40,000 apiece, which Wall Street deems conservative.
这场开发者大会的目的便是给出一个简单的答案：老黄已经无人能挡。一场全新的工业革命正在萌芽。黄仁勋将英伟达定位为建设面向未来的“AI工厂”的先锋。大模型工程师钟爱英伟达的图形处理器（GPU），对它的需求永无止境，以至于这种芯片目前已是供不应求。但没关系，英伟达宣布将于今年晚些时候将推出新一代最强AI加速卡Blackwell，其性能比自家已上市的GPU强出数倍，有望训练出更大、更聪慧的AI。黄仁勋表示，AI热潮之下，去年全球范围内投入数据中心的资金多达2500亿美元，并且还在以每年20%的速度增长。他的公司打算吞下大部分的“增长蛋糕”。为了将竞争对手远远甩在身后，英伟达将把Blackwell加速卡的价格定在3-4万美元。华尔街认为这一定价较为保守。
注释：
上周二，英伟达发布新一代Blackwell GPU架构，不仅刻意弱化了单芯片的存在感，而且没有明确GPU的代号，而是隐晦地称作“Blackwell GPU”。根据现有信息，全新Blackwell GPU……已知的芯片款式有3类——B100、B200、GB200超级芯片。
source: https://new.qq.com/rain/a/20240324A04CKZ00
In order to reap the fruits of this “accelerated-computing”, Nvidia wants to vastly expand its customer base. Currently the big users of its GPUs are the cloud-computing giants, such as Alphabet, Amazon and Microsoft, as well as builders of gen-ai models, such as Openai, maker of ChatGPT. But Nvidia sees great opportunity in demand from firms across all industries: health care, retail, manufacturing, you name it. It believes that many businesses will soon move on from toying with ChatGPT to deploying their own gen ais. For that, Nvidia will provide self-contained software packages that can either be acquired off the shelf or tailored to a user’s needs. It calls them nims, Nvidia Inference Microservices. Crucially, they will rely on (mostly rented) Nvidia GPUs, further tying customers into the firm’s hardware-software ecosystem.
为了收获“加速计算”浪潮的硕果，英伟达想要大幅扩大客户群体。目前，其GPU的主要用户是云计算巨头，如Alphabet、亚马逊和微软，以及构建生成式AI模型的企业，如发布ChatGPT的OpenAI。在医疗保健、零售业、制造业等各行各业的需求中，英伟达都洞察到了绝佳机会。它相信，许多企业很快就会从摆弄ChatGPT转向部署自己的生成式AI。为此，英伟达将提供独立的软件包，既有现成的产品，也可以根据用户的需求来定制。英伟达称它们为NIMs（即Nvidia Inference Microservices，英伟达推理微服务）。关键之处在于，NIMs的部署将依赖于（大部分为租用）英伟达的GPU，从而将客户牢牢拴在英伟达的软硬件生态之中。
注释：
1. NVIDIA 的 AI 平台适用于任何高瞻远瞩的企业使用，而且比以往任何时候都更加易于应用。NVIDIA AI Enterprise 5.0于3月18日发布，其中包括NVIDIA微服务、用于部署生成式AI应用和加速计算的可下载软件容器。领先的云服务提供商、系统构建商和软件供应商均可提供，并且Uber等客户已开始使用该产品。
source: https://blogs.nvidia.cn/2024/03/18/microservices-ai-enterprise/
2. NIM是一套经过优化的云原生微服务，旨在缩短上市时间，并简化生成式AI模型在云、数据中心和GPU加速工作站的任何位置的部署。
source:https://developer.nvidia.cn/zh-cn/blog/nvidia-nim-offers-optimized-inference-microservices-for-deploying-ai-models-at-scale/
So far, so star-spangled. But it is not all peace and love at Woodstock. You need only to recall the supply-chain problems of the pandemic, as well as the subsequent Sino-American chip wars, to see that dangers lurk. Nvidia’s current line-up of GPUs already faces upstream bottlenecks. South Korean makers of high-bandwidth memory chips used in Nvidia’s products cannot keep up with demand. tsmc, the world’s biggest semiconductor manufacturer, which actually churns out Nvidia chips, is struggling to make enough of the advanced packaging that binds GPUs and memory chips together. Moreover, Nvidia’s larger integrated systems contain around 600,000 components... That underscores the geopolitical risks if America’s tensions with its strategic rival keep mounting.
截至目前，一切成功，充满了鲜花与掌声。但伍德斯托克音乐节上也不全是爱与和平。只需回想一下疫情期间的供应链问题，你就能发现危机四伏。目前，英伟达的GPU产品线正面临上游瓶颈。其产品中使用的高带宽内存芯片由韩国公司生产，目前已供不应求。英伟达的芯片由全球最大的半导体制造商台积电（TSMC）代工，该公司的先进封装产能（用于将GPU和内存芯片结合在一起）同样不足。此外，英伟达的大型集成系统包含了约60万个组件……如果美国与其战略对手的关系持续紧张，那么地缘政治风险便更加凸显。
Troubles may lie downstream, too. The ai chips are energy-hungry and need plenty of cooling. There are growing fears of power shortages because of the strain that GPU-stuffed data centres will put on the grid. Mr Huang hopes to solve this problem by making GPUs more efficient. He says the mightiest Blackwell system, known pithily as gb200nvl72, can train a model larger than ChatGPT using about a quarter as much electrical power as the best available processors.
下游可能也同样面临麻烦。AI芯片非常耗能且需要长时间冷却。塞满了GPU的数据中心将给电网带来压力，因此人们愈发担心电力短缺问题。黄仁勋希望通过提高GPU的效率来解决这个问题。他表示，最强大的Blackwell系统（即GB200 NVL72）可以训练出一个比ChatGPT更大的模型，且能耗只有现下顶尖处理器的四分之一。
注释：
3月19日，英伟达召开了“2024GTC”大会，在会上重磅发布了新一代AI专用GPU GB200 NVL72。GB200NVL72是一款专门面向万亿参数大模型训练、推理的产品。与上一代的H100 GPU相比，GB200 NVL72可以将训练效率提升4倍，数据处理提升6倍，实时推理效率提升30倍。
source: https://new.qq.com/rain/a/20240319A05JSG00
But that is still almost 20 times more than pre-ai data-centre servers, notes Chase Lochmiller, boss of Crusoe Energy Systems, which provides low-carbon cloud services and has signed up to buy the gb200nvl72. And however energy-efficient they are, the bigger the GPUs, the better the ais trained using them are likely to be. This will stoke demand for ais and, by extension, for GPUs. In that way, as economists pointed out during a previous industrial revolution in the late 19th century, efficiency can raise power consumption rather than reduce it. “You can’t grow the supply of power anything like as fast as you can grow the supply of chips,” says Pierre Ferragu of New Street Research, a firm of analysts. In a sign of the times Amazon Web Services, the online retailer’s cloud division, this month bought a nuclear-powered data centre.
低碳云服务供应商克鲁索能源系统公司（Crusoe Energy Systems）已签约购买GB200 NVL72，但该公司老板查斯·洛克米勒（Chase Lochmiller）指出，与前AI时代的数据中心服务器相比，其耗电量仍然高出近20倍。而且能效高低不论，GPU越强大，所训练的AI能力就可能越强。这将刺激对AI的需求，进而刺激对GPU的需求。正如19世纪末第二次工业革命时经济学家所指出的，这样一来，能效的提升会增加而不是减少能源消耗。分析公司新街研究（New Street Research）的皮埃尔·费拉居（Pierre Ferragu）说:“你不可能像增加芯片供应一样快速地增加电力供应。”这个月，在线零售商亚马逊的云计算部门亚马逊网络服务（Amazon Web Services）收购了一个核动力数据中心，这多少说明了一些问题。
’Scuse me while I kiss AI
对不起，我要亲吻AI
Mr Huang is not blind to these risks, even as he dismisses the more typical concerns about gen ai—that it will destroy work or wipe out humanity. In his telling, the technology will end up boosting productivity, generating profits and creating jobs—all to the betterment of humankind. Hendrix famously believed music was the only way to change the world. For Mr Huang, it is a heady mix of science, engineering and maths.
AI摧毁就业、消灭人类这类更为广泛的担忧在老黄看来不过是杞人忧天，但他并不是没有意识到这些风险。按照他的说法，AI最终将提高生产力，创造利润，创造就业机会——所有这些都将造福人类。亨德里克斯坚信音乐是改变世界的唯一途径。对黄仁勋来说，集科学、工程和数学为一体的“上头”组合将最终改变世界。
3
观点|评论|思考
4
愿景
打造
01 2024全年写作精品课
微信扫码关注该文公众号作者