AMD与英伟达，谁更胜一筹？ | 经济学人商业
1
思维导图：
Angela，如果这纷乱的世界让我沮丧，我就去看看她们眼中的光芒。
2
Business | The AI-chip race
经济学人商业 | AI芯片竞赛
英文部分选自经济学人20240203期商业版块
Business | The AI-chip race
经济学人商业 | AI芯片竞赛
Could AMD break Nvidia’s chokehold on chips?
超威半导体公司（AMD）能否打破英伟达在芯片领域的垄断格局？
Taking on the top AI chipmaker will be hard—but maybe not impossible
挑战AI芯片巨无霸实属不易，但并非无望
“It is the most advanced ai accelerator in the industry,” boasted Lisa Su, boss of Advanced Micro Devices (AMD), at the launch in December of its new MI300 chip. Ms Su rattled off a series of technical specifications: 153bn transistors, 192 gigabytes of memory and 5.3 terabytes per second of memory bandwidth. That is, respectively, about 2, 2.4 and 1.6 times more than the H100, the top-of-the-line artificial-intelligence chip made by Nvidia. That rival chipmaker’s prowess in the semiconductors fuelling the ai boom has, over the past year, turned it into America’s fifth-most-valuable company, with a market capitalisation of $1.5trn. Yet most experts agreed that the numbers and Ms Su weren’t lying: the MI300 does indeed outshine the H100. Investors liked it, too—AMD’s share price jumped by 10% the next day.
去年12月的MI300芯片发布会上，AMD总裁苏姿丰（Lisa Su）夸耀道：“这是业内最先进的AI加速器。”她细数了该产品的一系列技术规格：1530亿个晶体管、192GB内存、每秒5.3TB的内存带宽，分别达到英伟达顶尖AI芯片H100的2倍、2.4倍和1.6倍。过去的一年里，英伟达在半导体领域的非凡实力助推了AI繁荣，并让自己成为了美国第五大上市公司，市值达1.5万亿美元。然而，大部分专家都认为，苏姿丰所言非虚：MI300确实比H100更胜一筹。AMD也赢得了投资者们的青睐，发布会第二天，AMD股价便飙升了10%。
On January 30th, in its quarterly earnings call, AMD announced that it expected to sell $3.5bn-worth of MI300s this year. It also reported strong revenues of $23bn in 2023, four times what they had been in 2014, when Ms Su became chief executive. Its market value is up 100-fold on her watch, to $270bn. Relative to forecast profits in the next 12 months, its valuation is richer even than Nvidia’s. Last year it displaced Intel, which once ruled American chipmaking, as the country’s second-most-valuable semiconductor company. Now it is taking aim at the biggest.
1月30日，在其2023年第四季度业绩电话会上，AMD预测今年MI300销售额将达35亿美元。财报还显示，2023年整体营收高达230亿美元，是2014年苏姿丰刚上任CEO时的四倍。在苏姿丰的领导下，AMD市值增加了100倍，达2700亿美元。以未来12个月预期利润这个口径来计算，AMD估值甚至比英伟达还高。去年，AMD就超过了曾一度称霸美国芯片制造业的英特尔，一跃成为美国市值第二高的半导体企业。如今，它剑指王位。
Such ambition would have seemed fanciful a decade ago. Back then, recalls Mark Papermaster, AMD’s technology chief, AMD was facing an “existential crisis”. In 2008 it had spun off its chip-fabrication business to focus on designing processors, outsourcing manufacturing to contract chipmakers such as tsmc of Taiwan. The idea was to be better able to compete on blueprints with Intel, whose vast fabrication capacity AMD could not hope to match.
如此雄心放在十年前仿佛是痴人说梦。回忆往昔，AMD的首席技术官马克·佩博马斯特（Mark Papermaster）说道，当时AMD正面临着一场“生存危机”。2008年，它剥离了芯片制造业务，专注于处理器设计，将生产外包给中国台湾的台积电公司（TSMC）等芯片代工制造商。此举是为了更好地与英特尔在产品设计上一较高下；毕竟，英特尔强大的制造能力让AMD难以望其项背。
It didn’t work. Several of AMD’s chips flopped. Sales of its central processing units (cpus), mostly for personal computers, were collapsing. In 2013 it sold and leased back its campus in Austin to raise cash. A year later Ms Su inherited a net-debt pile of more than $1bn, a net annual loss of $400m and a market value of less than $3bn, down from $20bn in 2006.
这一招并未奏效。AMD好几款芯片都一败涂地，主要应用于个人电脑的中央处理器（CPU）的销量也萎靡不振。2013年，为了筹措资金，AMD对其在奥斯汀的园区先卖后租。一年之后，苏姿丰接手了这家公司。当时AMD的净债务逾10亿美元，年净亏损达4亿美元，市值从2006年的200亿美元跌至不足30亿美元。
She realised that the only way for AMD to get back in the game was to steer it away from the sluggish pc market and focus on more promising areas like cpus for data-centre servers and graphics processing units (gpus, which make video-game visuals lifelike) for gaming consoles. She and Mr Papermaster took a gamble on a new cpu architecture designed to beat Intel not just on price, but also on performance.
她意识到，AMD要想重振雄风，唯一办法就是从低迷的个人电脑市场转战到更有前景的领域，比如数据中心服务器的CPU与游戏机的图形处理单元（即GPU，它可以让电子游戏的视觉效果栩栩如生）。苏姿丰和佩博马斯特放手一搏，设计了全新的CPU架构，力图在价格和性能上实现对英特尔的双杀。
When the going got tough
逆境之中
The idea was to use a Lego-like approach to chip building. By breaking a chip up into smaller parts, AMD could mix and match blocks to assemble different types of chip, at a lower cost. When the first such composite chips were released in 2017, they were zippier and cheaper than rival offerings from Intel, possibly in part because Intel was distracted by its own problems (notably repeated manufacturing slip-ups as it moved to ever tinier transistors). In the past ten years AMD’s market share in lucrative server cpus has gone from nothing to 30%, breaking Intel’s monopoly.
其思路是采用类似乐高积木的方法来构建芯片。把芯片拆解成小模块，就可以用较低的成本，混搭不同小模块，组成不同类型的芯片。2017年，AMD首次推出这种复合芯片，它们比英特尔的芯片性能更好，而且还要更便宜（当然，部分原因可能是英特尔自己走了弯路：一心追求更小的制程，却在生产时屡出差错）。过去十年，AMD在服务器CPU这个利润丰厚的市场中所占份额从0增长到30%，打破了英特尔的垄断。
Having faced down one giant, AMD now confronts another. The contest with Nvidia is different. For one thing, it is personal—Ms Su and Jensen Huang, Nvidia’s Taiwanese-born boss, are distant relatives. In contrast to Intel, Nvidia is, like AMD, a chip designer and thus less prone to production missteps. More importantly, the stakes are higher. Nvidia’s market value of $1.5trn is predicated on its dominance of the market for gpus—not because of their usefulness in gaming but because they also happen to be the best type of chip to train ai models. Ms Su expects global sales of ai chips to reach $400bn by 2027, up from perhaps $40bn last year. Does she stand a chance against Nvidia?
和英特尔一较高下后，AMD又开始挑战另一巨头——英伟达。这次的较量有所不同。首先，这事关两家公司的掌舵者个人——AMD的苏姿丰和英伟达的黄仁勋是远亲。此外，与英特尔不同，英伟达也是一家芯片设计公司，因此不太可能出现生产失误。更重要的是，此次竞争的风险更高。支撑起英伟达1.5万亿美元市值的，是其在GPU市场中的主导地位。而这并不是因为GPU对电脑游戏至关重要，而是因为GPU恰好是训练人工智能模型的最佳芯片类型。苏姿丰预计，到2027年，全球人工智能芯片的销售额将从去年的约400亿美元增长至4000亿美元。她有希望战胜英伟达吗？
Nvidia is a formidable rival. Both its revenues and operating margins are nearly three times AMD’s. According to Jefferies, an investment bank, the company dominates the market for ai accelerator chips, accounting for 86% of such components sold globally; before the launch of the MI300, AMD barely registered. Nvidia also offers network gear that connects clusters of chips, and software, known as cuda, to manage ai workloads. Nvidia has dominated ai chipmaking because it has offered the best chips, the best networking kit and the best software, notes Doug O’Laughlin of Fabricated Knowledge, a research firm.
英伟达可不是省油的灯，无论是收入还是经营利润，几乎都是AMD的三倍。投资银行Jefferies称，英伟达占据了全球AI加速器芯片市场86%的份额；而AMD在发布MI300前几乎没有市场份额。同时，英伟达还能提供可以连接芯片集群的网络设备，以及管理AI工作负载的CUDA软件。Fabricated Knowledge研究公司的道格·奥劳夫林（Doug O’Laughlin）指出，英伟达凭借其最强大的芯片、最强大的网络设备和最强大的软件主导了整个AI芯片制造业。
AMD’s new processor shows it can compete with Nvidia on semiconductor hardware. This, Mr Papermaster says, is the result of a ten-year investment. AMD is spending nearly $6bn a year on research and development, nearly as much as its larger rival—and twice as much as a share of sales (see table). This has enabled it to adapt its Lego approach to gpus. Combining a dozen blocks—or “chiplets”—into a single chip lets AMD put processors and memory close to each other, which boosts processing speed. In December Openai, maker of Chatgpt and the world’s hottest ai startup, said it would use the MI300s for some of its training.
AMD的新处理器表明，它在硬件上的实力足以与英伟达竞争，佩博马斯特认为这是十年来投资的成果。AMD规模不如英伟达，但每年在研发上投入近60亿美元，跟英伟达不相上下；如果以占销售额的比重来看，AMD对研发的投入程度是英伟达的两倍（见上表）。这使得AMD能将搭积木的方式应用到GPU上——将十来个模块（即“芯片块”）组合成单个芯片，能让处理器和内存在物理上挨得更近，从而提高处理速度。去年12月，ChatGPT的创造者、全球最热门的AI初创公司OpenAI表示将在部分模型训练中使用MI300。
To outdo Nvidia on networking and software, AMD is teaming up with other firms. In December it announced a partnership with makers of networking gear, including the two largest, Broadcom and Cisco. It is also supporting an open-source initiative for chip-to-chip communication called Ultra Ethernet Consortium as an alternative to InfiniBand, a rival championed by Nvidia.
为了能在网络设备和软件上超过英伟达，AMD正与其他公司携手合作。去年12月，AMD宣布与博通（Broadcom）和思科（Cisco）等网络设备制造商合作。AMD同时还支持一项名为“超级以太网联盟”（Ultra Ethernet Consortium）的芯片间通信开源倡议，对标由英伟达倡导的“无限带宽”技术（InfiniBand）。
Chomping at the byte
步步紧逼
Nvidia’s lead in software will be harder to close. It has been investing in cuda since the mid-2000s, well before the current ai wave. ai developers and researchers love the platform, which allows them to fine-tune the performance of Nvidia processors. AMD hopes to tempt customers away from Nvidia by making its software, rocm, open-source and providing tools to make the switch smoother, by translating cuda programs into rocm ones.
英伟达在软件领域的领先地位将更难撼动。英伟达从2000年代中期就开始投资CUDA，比当下的AI浪潮要早多了。AI开发者和研究者对CUDA钟爱有加，他们可以通过这一平台对英伟达处理器的性能进行微调。为了将客户从英伟达的平台吸引过来，AMD将开源其软件ROCM并提供编译工具，使得从CUDA转战ROCM的过程更加顺畅。
注释：chomp: If a person or animal chomps their way through food or chomps on food, they chew it noisily. (人、动物进食时)大声地咬，嚼得很响
Beating Nvidia at its own game will not be easy. Mr Huang’s firm is not standing still. It recently announced plans to bring out a new chip every year instead of every two years. The tech giants with the grandest ai ambitions—Alphabet, Amazon, Meta and Microsoft—are busily designing their own accelerator chips. Despite AMD’s robust sales, investors were disappointed with its forecast for MI300 shipments. Its share price dipped by 3% the day after it reported its latest results.
在英伟达的地盘打败英伟达绝非易事。黄仁勋的公司并没有停滞不前。近期，英伟达宣布将不再每两年推出一款新芯片，而是计划每年推出一款。Alphabet、亚马逊、Meta和微软等在AI领域有着鸿鹄之志的科技巨头都忙着设计自己的加速器芯片。尽管AMD销量强劲，投资者对其MI300的预期出货量仍然感到失望，其股价在公布最新财报的第二天就下跌了3％。
注释：
1.beat (one) at (one's) own game: to defeat or do better than someone in an activity which they have chosen or in which they think they are strong赢某人的看家本领；打败某人的强项
2.dip: If the amount or level of something dips, it becomes smaller or lower, usually only for a short period of time. (数量或水平，通常短暂地)下降，下跌，减少
Still, AMD has one big thing going for it. It is not Nvidia. ai companies are desperate for an alternative to its larger rival, whose dominant position allows it to charge steep prices and, with demand outstripping supply, ration chips to buyers. Despite efforts to design their own hardware, big tech firms will rely on chipmakers for a while, and AMD gives them options, notes Vivek Arya of Bank of America. Microsoft and Meta have already announced plans to use AMD’s gpus in their data centres. And if Nvidia slips up, AMD will be there to pick up the Lego pieces. Just ask Intel.
尽管如此，AMD仍有一项巨大优势：AMD不是英伟达。AI企业不愿见到英伟达一手遮天——它靠着市场统治地位疯狂涨价，并在需求旺盛时配额销售。美国银行（Bank of America）的维维耶·阿里亚（Vivek Arya）表示，尽管大型科技公司仍在努力设计自己的硬件，但在一段时间内，它们仍需要依赖芯片制造商，而AMD可以使它们多一种选择。微软和Meta已经宣布计划在自己的数据中心使用AMD的GPU。如果英伟达出现失误，AMD将依仗其“乐高大法”捡漏。不信去问问英特尔。
注释：
1.If you have something going for you, or if something has something going for it, there are certain advantages that will make the end result successful有优势
例：They've got a happy marriage, great careers, wonderful kids - in fact they've got everything going for them.
2.steep: If you say that the price of something is steep, you mean that it is expensive. (价格)过高的，昂贵的
3.outstrip: If one thing outstrips another, the first thing becomes larger in amount, or more successful or important, than the second thing. 超过；胜过
4.ration: When something is rationed by a person or government, you are only allowed to have a limited amount of it, usually because there is not enough of it.（通常指由于紧缺）配给，定量供应
5.The phrase "pick up the Lego pieces" is used metaphorically to describe dealing with failure, chaos, or the task of rebuilding after something has fallen apart or gone wrong. It's a visual image that compares a messy situation to the scattered pieces of a Lego project that needs to be put back together.
翻译组：
Ryan，学海无涯，译无止境
Charlie，剃刀边缘无比锋利
Alfredo ，小小少年，很多烦恼
校对组：
Shulin，时间是个好东西
Cecilia，今年生活得贵妇一点
3
观点|评论|思考
本期感想
Neil，男，外贸民工，经济学人铁粉
3
愿景
打造
01 第一期口译打卡营
团队由毕业于蒙特雷，纽卡，北外，北二外等小伙伴组成
有丰富口译实战经验，课程内容包括
听辨交传笔记法、短记复述练习、视译专题等
助力大家拿下三口、顺利开启口译学习！
点击下图，即可了解口译课详情！
微信扫码关注该文公众号作者