The end of Moore’s law will not slow the pace of change
摩尔定律的终结不会减缓变革的步伐

来源The Economist, 词数890
来源The Economist， 词数890

The chipmaking industry has always existed in a state of paranoid optimism. Ever since Gordon Moore’s observation that processing power would double roughly every two years was encoded by others into “Moore’s law”, a chorus of experts has been warning of its end. That dread is tinged with a fierce belief that there is always a way to stave off the inevitable. The results have been nothing short of spectacular. In the past 50 years, processors have come to operate tens of thousands of times faster, and store a million times more data in the same area. The cost of a transistor has also fallen by a factor of a billion, making technology a global deflationary force. They are also ubiquitous: semiconductors are now the third-most traded commodity in the world by value, after oil and cars.
芯片制造行业一直处于一种偏执乐观的状态。自从 Gordon Moore 观察到处理能力大约每两年翻一番被编入“摩尔定律”以来，专家们一直在警告它的终结。这种恐惧中夹杂着一种强烈的信念，即总有办法避免不可避免的事情。结果简直是惊人的。在过去 50 年中，处理器的运行速度提高了数万倍，在同一区域存储的数据增加了一百万倍。晶体管的成本也下降了 10 亿倍，使技术成为全球通货紧缩的力量。它们也无处不在：按价值计算，半导体现在是世界上交易量第三大的商品，仅次于石油和汽车。

But decades of success have not quelled the industry’s jitters over the end of Moore’s maxim. In 2017 Jensen Huang, boss of Nvidia, now the world’s most valuable chip company, declared that the law was dead. In June, Pat Gelsinger, chief executive of Intel, the firm that Mr. Moore co-founded, insisted that the law was “alive and well”. This Technology Quarterly has argued that if not dead, Moore’s law is on life support.
但几十年的成功并没有平息该行业对摩尔格言结尾的不安。2017 年，英伟达（Nvidia）的老板黄仁勋（Jensen Huang）宣布该法律已死。今年 6 月，英特尔首席执行官帕特·基辛格 （Pat Gelsinger） 先生表示。 摩尔是该法律的联合创始人，他坚称该法律“生机勃勃”。本期《技术季刊》认为，如果摩尔定律没有死，那么它就是关于生命维持的。

For much of the transistor’s history it followed a “happy scaling” path—as the logic gates shrank, they got faster and used less power. That era is over. Leading-edge AI processors cram more transistors on a single chip or stack multiple “chiplets” into one package to boost computing oomph. But the performance gains have come at a cost: the energy consumed by a chip has ballooned. A single Blackwell chip, Nvidia’s latest, runs five times faster than its predecessor, but uses 70% more power in the process.
在晶体管历史的大部分时间里，它都遵循着一条“快乐扩展”的道路——随着逻辑门的缩小，它们会变得更快，消耗的功率也更少。那个时代已经结束了。领先的 AI 处理器在单个芯片上塞入更多晶体管，或将多个“小芯片”堆叠到一个封装中，以提高计算能力。但性能提升是有代价的：芯片消耗的能量激增。单个 Blackwell 芯片是 Nvidia 的最新芯片，运行速度比其前身快 5 倍，但在此过程中消耗的功率增加了 70%。

Data centers lash hundreds or thousands of these power-hungry chips together to run large artificial-intelligence (AI) models. By some estimates, Open-AI, maker of ChatGPT, guzzled more than 50 gigawatt-hours of electricity to train its latest model. The International Energy Agency calculates that in 2022 data centers consumed 460 terawatt-hours, or almost 2% of global electricity demand. The agency expects this figure to double by 2026.
数据中心将成百上千个这些耗电的芯片捆绑在一起，以运行大型人工智能 （AI） 模型。据估计，Chat GPT 的制造商 Open-AI 消耗了超过 50 吉瓦时的电力来训练其最新模型。国际能源署计算出，2022 年数据中心消耗了 460 太瓦时，占全球电力需求的近 2%。该机构预计，到 2026 年，这一数字将翻一番。

The tricks chipmakers have used to boost processing power for AI models without causing runaway energy growth hint at shifts in the semiconductor industry. The first change is the decline of the computer as a general-purpose machine. Neil Thompson of MIT argues that, for decades, Moore’s law held computing together. Every successive generation of semiconductor technology was faster and more energy efficient than the previous one. That allowed the tech world to rely on a universal processor—the central processing unit (CPU)—that could be programmed for lots of tasks. But the end of Moore’s law makes it harder now to improve performance across all applications.
芯片制造商用来提高 AI 模型的处理能力而不导致能源失控增长的技巧暗示了半导体行业的转变。第一个变化是计算机作为通用机器的衰落。麻省理工学院的 Neil Thompson 认为，几十年来，摩尔定律将计算联系在一起。每一代半导体技术都比上一代更快、更节能。这使得科技界能够依赖一个通用处理器，即中央处理器 （CPU），它可以针对许多任务进行编程。但是摩尔定律的结束使得现在提高所有应用程序的性能变得更加困难。

The response, in the case of AI chips, has been to specialize or fine-tune chips for specific software. Mr. Thompson believes this could split computing into two lanes: the fast lane, where cutting-edge applications benefit from powerful customized chips, and the slow lane, where ordinary applications get stuck using general-purpose chips whose progress is slowing.
就 AI 芯片而言，回应 是针对特定软件专门化或微调芯片。先生 Thompson 认为，这可能会将计算分为两条车道：快车道，尖端应用程序受益于强大的定制芯片，以及慢速车道，普通应用程序使用进展缓慢的通用芯片。

The need to specialize has precipitated a second shift—the rise of firms that control both hardware and software. For over five decades, the tech world was neatly split into two camps: the hardware crowd who tinkered with their circuits and the coding geeks who wrote software. The “Wintel” alliance whereby computers would run Microsoft’s Windows operating system on chips made by Intel has been one of the most successful partnerships in technology history.
对专业化的需求促成了第二个转变——同时控制硬件和软件的公司崛起。五十多年来，科技界整齐地分为两个阵营：修补电路的硬件人群和编写软件的编码极客。“Wintel”联盟让计算机在英特尔制造的芯片上运行 Microsoft 的 Windows 操作系统，这是技术史上最成功的合作伙伴关系之一。

Now the wall separating the two camps has cracked. Bill Dally, Nvidia’s chief scientist, says that improvements in software and chip architecture yield bigger gains than moving to newer manufacturing processes. At the cutting edge, specialist silicon is the future and the giants are doing a lot of it themselves.
现在，分隔两个阵营的墙已经破裂。Nvidia 的首席科学家 Bill Dally 表示，与转向更新的制造工艺相比，软件和芯片架构的改进会产生更大的收益。在最前沿，特种硅是未来，巨头们自己也在做很多事情。

Apple, Amazon, Google, Microsoft and Meta all now use custom silicon that is optimized for their own software. Google’s processors are designed to run TensorFlow, its machine-learning software. Apple’s homemade chips are tuned to run its own software on the gadgets it makes. These firms partner with firms like Broadcom, an American chip company, to design these chips, and a foundry like TSMC to build them. Nvidia is the only one to have made a great business out of making AI chips for others—but this is in part because its chips are optimized for CUDA, its software platform, which has become an industry standard.
Apple、Amazon、Google、Microsoft 和 Meta 现在都使用针对自己的软件进行了优化的定制芯片。Google 的处理器旨在运行其机器学习软件 TensorFlow。Apple 的自制芯片经过调整，可以在其生产的小工具上运行自己的软件。这些公司与美国芯片公司 Broadcom 等公司合作设计这些芯片，并与 TSMC 等代工厂合作制造这些芯片 。Nvidia 是唯一一家通过为他人制造 AI 芯片而取得巨大成就的公司，但这部分原因是其芯片针对其软件平台 CUDA 进行了优化，CUDA 已成为行业标准。

According to BCG，a consultancy, between 2018 and 2023 the “big six” firms (the five tech giants plus Nvidia) accounted for roughly 60% of the value in the technology sector. But that will not slow innovation. This new world is even more inventive than the previous era that sparked the computer revolution. There are now so many more ways for computers to become better—through silicon cleverness, better packaging, innovative chip designs and new materials. It will be competitive in ways the Wintel world never could be, and may create things even more wonderful. A combination of silicon, dread and fierce belief are going to achieve even more in the next 70 years than they have in the past 70.
根据咨询公司 BCG 的数据，2018 年至 2023 年期间，“六大”公司（五大科技巨头加上英伟达）约占科技行业价值的 60%。但这不会减缓创新。这个新世界比引发计算机革命的上一个时代更具创造性。现在，计算机有更多方法可以变得更好 - 通过硅的智能、更好的封装、创新的芯片设计和新材料。它将以 Wintel 世界前所未有的方式具有竞争力，并且可能会创造更加精彩的事情。硅、恐惧和强烈信念的结合将在未来 70 年取得比过去 70 年更大的成就。