内容来源：https://www.wired.com/story/huawei-chip-queen-moores-law-tau/

内容总结：

华为“芯片女皇”何庭波宣布新路径：2026年冬季前推出突破性芯片，有望缩小中西技术差距

在备受全球关注的AI芯片竞赛中，华为传来新动向。华为海思总裁何庭波近日在上海举行的IEEE国际电路与系统研讨会上表示，其工程团队开发出了一种全新的芯片优化方法，并自信地认为，这一“新路径”将在未来几年内缩小中国芯片与西方芯片的性能差距。

何庭波被誉为华为的“芯片女皇”。她指出，华为的新方法不再单纯追求在单一硅片上集成更多晶体管，而是聚焦于加速芯片、电路乃至整个计算系统的运行速度。她将这一新法则命名为“Tau’s Scaling Law”，并直言其已取代摩尔定律，成为海思的新指导原则。

“我们找到了一条新路，”何庭波在大会上承诺，华为将在未来几个月内通过发布新芯片来证明这一方法的可行性。她更是放出豪言：“2026年冬季之前，我们会带来惊喜。这不是饱和，也不是延续，而是一次巨大的飞跃。”

当前美国出口管制切断了华为与台积电的合作，使其只能依赖使用较老一代光刻机的中芯国际。业界普遍认为，中国在尖端AI开发方面至少落后西方五年。然而，芯片行业本身也已遭遇摩尔定律的物理极限，当晶体管缩小至几纳米时，量子效应会干扰其正常运行。

华为的宣布表明，公司已找到绕过这些物理限制的方法。何庭波强调，六年前几何微缩（指光刻小型化）就已经达到平台期，他们意识到半导体演进远不止于此。她列举了包括“LogicFolding”在内的多项技术，该技术能减少电路内部关键逻辑运算所需的时间。

此外，海思还通过考虑纳米级电子现象、优化芯片协同设计以及开发能加速芯片间通信的互连技术来提升性能。何庭波解释道：“无论是AI训练还是推理，赢点不仅在于缩短计算时间，更在于缩短数据在芯片间和芯片内部移动的时间。”

华为宣称，通过这一新方法，其将在2031年前生产出相当于1.4纳米制程工艺性能的组件。考虑到台积电预计在2028年引入该工艺，这相当于大幅缩短了中国芯片制造的落后差距。

独立半导体与AI政策分析师Lennart Heim评论认为，华为的策略表明其在单纯靠缩小和密集化芯片来获取性能提升方面遇到了瓶颈，因此正越来越多地依赖混合键合和3D芯片堆叠等技术。但何庭波信心十足地总结道：“这些创新将进入大规模生产，也许不是今年，但从2027年开始会持续实现。”

中文翻译：

人工智能芯片竞赛的剧情愈发扑朔迷离。
华为芯片设计子公司海思总裁何庭波表示，其工程师团队已研发出一种全新的半导体优化方法，她相信这一技术将在未来几年内缩小中国芯片与西方芯片的性能差距。
简而言之，华为的方法侧重于加速芯片、电路乃至整个计算系统的运算速度，而非将更多元件挤压到单一硅片上。
“我们找到了一条新路径，”何庭波上周在上海举行的电气与电子工程师协会电路与系统国际研讨会上表示。这位被誉为中国“芯片女皇”的何庭波承诺，华为将在未来数月内（预计通过新款芯片）证明这一新方案的可行性。
“2026年冬季之前，我们将带来惊喜，”她说，“不是饱和，不是延续，而是重大飞跃。”
何庭波将这一新方法命名为“陶氏缩放定律”，并称其已取代摩尔定律成为海思的指导原则。摩尔定律以英特尔联合创始人戈登·摩尔命名，其核心观点是计算能力的提升取决于芯片中晶体管或逻辑门数量约每两年翻一番。
目前制造尖端芯片需要借助价值数十亿美元的光刻设备、精密脆弱的供应链组件以及深厚工程专业知识，将元件蚀刻在硅片上。
美国出口管制禁止华为与全球领先芯片代工厂台积电合作。华为只能转而依赖使用较老一代光刻机的中国中芯国际。关键在于，这些限制削弱了中国利用自产芯片开发前沿人工智能的能力——据估算，中国在这一领域落后领先水平超过五年。
但芯片行业已开始触及摩尔定律的天花板。当晶体管宽度仅达几纳米时，量子效应会干扰其正常运行。许多芯片已在采用变通方案：例如苹果最强大的处理器，便是通过拼接两颗芯片形成性能更强的单一芯片。
华为的声明表明，该公司相信已找到突破这些限制的方法。这也意味着，旨在遏制中国芯片产业的制裁反而催生了创新，假以时日，这些创新或可让中国建立更先进的自主芯片产业，并与西方竞争。最终，华为等公司的创新可能削弱美国的技术优势。
“六年前，几何缩放在我们这里达到了极限，”何庭波周末表示，她指的是光刻技术微缩化，“我们很快意识到，半导体演进远不止几何缩放。”
她着重介绍了公司通过新方法提升芯片性能的多种途径。其中包括名为“逻辑折叠”的技术，该技术可缩短电路内执行关键逻辑运算所需的时间。
海思表示，其还在通过以下方式提升芯片性能：考量纳米级电子现象、优化芯片协同设计、开发加速芯片间通信的互连技术——这是训练大型人工智能模型的关键诀窍。
“无论是人工智能训练还是推理，制胜关键不仅在于缩短计算时间，更在于减少数据在芯片内部及芯片间的移动时间，”她说。
华为表示，将利用新方法在2031年前生产性能相当于1.4纳米制程工艺的组件。这将大幅缩短中国芯片制造的落后差距——台积电预计于2028年推出采用该制程的芯片。
何庭波的声明并不意味着华为已找到破解美国制裁的明确路径，且并非所有人都相信这一方案可行。独立半导体与人工智能政策分析师伦纳特·海姆指出，华为的策略表明该公司在单纯通过缩小芯片尺寸、提升密度来压榨性能方面已触及瓶颈。他认为，华为正越来越多地依赖混合键合、3D芯片堆叠等技术来提升性能。
但华为芯片女皇似乎坚信公司能够改变游戏规则。“这些创新将进入量产，”她在演讲中表示，“或许不是今年，但从2027年及以后将逐步实现。”
本文是威尔·奈特《人工智能实验室》新闻通讯的一期内容。此前通讯可查阅往期文章。

英文来源：

The plot thickens in the great AI chip race.
Tingbo He, president of Huawei’s chip-design subsidiary HiSilicon, says her company’s engineers have developed a novel way to optimize semiconductors—and she believes it will close the performance gap between Chinese and Western chips over the next few years.
Huawei’s method, in short, focuses on speeding up computations across chips, circuits, and entire computing systems, rather than squeezing ever-more components onto a single piece of silicon.
“We found a new path,” He said at the IEEE International Symposium on Circuits and Systems in Shanghai last weekend. He, who is known in China as Huawei’s “chip queen,” promised that the company would prove the viability of the new approach, presumably with a new chip, in the coming months.
“Before winter 2026, we will bring the surprise,” He said. “Not saturation, not continuation, but a big leap ahead.”
The chip queen calls the new approach Tau’s Scaling Law, and says it has replaced Moore’s Law as HiSilicon’s guiding principle. Moore’s Law, named for the Intel cofounder Gordon Moore, dictates that progress in computing depends on roughly doubling the number of transistors, or logic gates, packed into a chip every two years.
Minting cutting-edge chips currently involves etching components into silicon using billion-dollar lithographic equipment, a supply chain of exquisitely delicate components, and extensive engineering know-how.
US export controls prohibit Huawei from working with Taiwan Semiconductor Manufacturing Company (TSMC), the world's leading chip foundry. Huawei must instead rely on China’s SMIC, which uses an older generation of lithography machines. Crucially, restrictions limit China’s ability to develop frontier artificial intelligence using its own silicon. By some estimates, it’s more than five years behind the leading edge.
But the chip industry has begun running into the limits of Moore’s Law. When transistors are just a few nanometers wide, quantum effects interfere with their normal functioning. Many chips are already made with workarounds: Apple’s most powerful processors, for example, are built by stitching two chips together to make a more powerful single one.
Huawei’s announcement suggests that the company believes it has found a way around these limits. It also suggests that the sanctions aimed at kneecapping China’s chip industry have spurred innovations that may, over time, allow the country to build a more advanced domestic chip industry and compete with the West. In the end, innovations from companies like Huawei could erode America’s technological edge.
“Six years ago geometric scaling plateaued for us,” He said over the weekend, referring to lithographic miniaturization. “We soon realized semiconductor evolution is more than geometric scaling.”
He highlighted several ways that the company has advanced chip performance using its new approach. These include something called LogicFolding, which reduces the time required to perform key logical operations within a circuit.
HiSilicon says it’s also improving chip performance by accounting for nanoscale electronic phenomena; designing chips to work well together; and developing interconnects that speed chip-to-chip communication, a key trick for training large AI models.
“For both [AI] training and inference, the win is not just in shortening compute time. It is in shortening the time that data spends moving, between chips and inside a chip,” she said.
Huawei says it will use its new approach to produce components with performance equivalent to a 1.4-nanometer chipmaking process by 2031. This would amount to a significant reduction in China's chipmaking lag since TSMC is expected to introduce chips using this process in 2028.
He’s announcement doesn’t mean that Huawei has a clear path to defeating US sanctions, and not everyone is convinced it will be viable. Lennart Heim, an independent semiconductor and AI policy analyst, says Huawei’s strategy suggests the company is running into limits on how much more performance it can squeeze out by shrinking and densifying chips alone. Instead, he says, Huawei is increasingly relying on techniques like hybrid bonding and 3D chip stacking to improve performance.
But Huawei’s chip queen seems confident that the company will change the game. “These innovations will enter mass production,” she said in her speech. “Maybe not this year, but from 2027 and beyond.
This is an edition of Will Knight’s AI Lab newsletter. Read previous newsletters here.