英伟达黄仁勋谈超摩尔定律与 AI 计算的未来

Nvidia's Jensen Huang on Hyper Moore's Law and the Future of AI Computing

黄仁勋 Jensen Huang · No Priors 播客 · 2024-11-07 · 约 37 分钟 · 原视频 ↗

打开互动全文版（中英对照 + 朗读 + 问答）→

本期速览 · Overview

黄仁勋探讨英伟达的十年赌注，包括超摩尔定律缩放、数据中心级计算，以及平衡 AI 推理低延迟与高吞吐量的挑战。

Jensen Huang discusses Nvidia's 10-year bets, including hyper Moore's Law scaling, data center-scale computing, and the challenges of balancing low latency with high throughput for AI inference.

要点 · TL;DR

• 我们正处于超摩尔定律的起点，性能每年翻倍。
  We are at the start of hyper Moore's law, doubling performance yearly.
• 数据中心是新的计算单元，需整体优化。
  The data center is the new unit of computing, optimized as a whole.
• AI 将在几年内改变所有科学与工程领域。
  AI will transform all science and engineering within a few years.

核心观点 · Key points

• 我们正处于超摩尔定律曲线的起点，每年规模性能翻倍或增长两倍。
  We are at the beginning of a hyper Moore's law curve, doubling or tripling performance every year at scale.
• 跨算法、系统和数据中心的协同设计是持续 Scaling 的关键。
  Codesign across algorithm, system, and data center is key to continued scaling.
• 计算的新单位是数据中心，必须作为一个整体来构建和优化。
  The new unit of computing is the data center, which must be built and optimized as a whole.
• AI 将在几年内改变所有科学与工程领域。
  AI will transform every field of science and engineering within a few years.
• 英伟达的架构兼容性确保软件随处运行，保护开发者投资。
  Nvidia's architectural compatibility ensures software runs everywhere, protecting developer investment.

反共识 · Contrarian takes

• 推理时间 Scaling 需要低延迟和高吞吐量，这两者相互矛盾。
  Inference time scaling requires both low latency and high throughput, which are at odds.
• 新型数据中心不是多租户文件存储，而是生产智能的单租户 Token 工厂。
  New data centers are not multi-tenant file stores but single-tenant token factories producing intelligence.
• SaaS 平台不会被颠覆；它们将因上层专门的 AI 智能体而蓬勃发展。
  SaaS platforms are not disrupted; they will flourish with specialized AI agents on top.
• 小型专用模型可以在狭窄领域达到超人性能。
  Small, specialized models can achieve superhuman performance in narrow domains.
• AI 芯片设计已不可或缺；没有它，英伟达无法构建 Hopper。
  AI chip design is already essential; Nvidia couldn't build Hopper without it.
• 计算边际成本在 10 年内下降了百万倍，使得穷举式软件生成成为可能。
  The marginal cost of computing has dropped a million times in 10 years, enabling exhaustive software generation.

本期章节 · Chapters（共 10）

• 引言与重大赌注 Introduction and Big Bets
• 优化与基础设施 Optimization and Infrastructure
• 训练与推理基础设施 Infrastructure for Training and Inference
• 软件经济学与英伟达战略 Software Economics and Nvidia's Strategy
• 英伟达架构兼容性优势 Nvidia's architectural compatibility and its benefits
• 芯片设计优化与 AI Chip design optimization and AI
• 具身智能与机器人 Embodiment and robotics
• AI 员工与企业未来 AI employees and company future
• AI 中被低估的领域 Underappreciated areas in AI
• 结束语 Closing Remarks

阅读全文双语转录 →