Cerebras 联合创始人兼 CEO Andrew Feldman 讨论为什么大芯片是加速 AI 处理的最佳方式，以及他的公司如何制造出计算史上最大的芯片。
Andrew Feldman, co-founder and CEO of Cerebras, discusses why big chips are the best way to process AI faster, and how the company built the largest chip in computer history.
要点 · TL;DR
每秒每用户 token 数成为 AI 生产力的核心指标，推理速度是新的决胜点。 Tokens per second per user is the key measure of AI productivity, making inference speed the new battleground.
晶圆级芯片以巨大 SRAM 绕开 HBM 瓶颈，推理速度远超 GPU。 Wafer-scale chips with huge SRAM sidestep the HBM bottleneck, delivering far faster inference than GPUs.
真正的瓶颈是存储、封装与代工产能；AI 需求真实且供不应求。 Real bottlenecks are memory, packaging, and fab capacity; AI demand is real and outstripping supply.
核心观点 · Key points
推理速度——每秒每用户令牌数——现在是AI生产力的主导指标。 Inference speed—tokens per second per user—is now the dominant metric for AI productivity.
带有海量SRAM的晶圆级芯片减少了权重移动时间，使其在推理上远快于GPU。 Wafer-scale chips with massive SRAM cut weight-moving time, making them far faster for inference than GPUs.
主要瓶颈是HBM内存、CoWoS封装和3纳米工厂产能——不仅仅是算力。 Major bottlenecks are HBM memory, CoWoS packaging, and 3nm fab capacity—not just compute.
AI需求真实且超过供给；行业在追逐已有需求，而非建设投机性产能。 AI demand is real and outstrips supply; the industry is chasing existing demand, not building speculative capacity.
中国是工业对手，在电网投资和开源模型上很强，但在芯片上落后。 China is an industrial adversary with grid investments and strong open-source models, though behind in chips.
CUDA的护城河正在侵蚀；领先模型现在在TPU和Trainium上训练，无需CUDA。 CUDA's moat is eroding; leading models now train on TPUs and Trainium without CUDA.
反共识 · Contrarian takes
GPU在推理上本质很慢，因为从HBM移动权重占主导；Cerebras大约快2500倍。 GPUs are fundamentally slow at inference because moving weights from HBM dominates; Cerebras is about 2500x faster.
完美的时机来自十年的坏时机——Cerebras在2020年解决了难题，但无人问津。 Perfect timing comes from horrible timing for a decade—Cerebras solved a hard problem in 2020 and nobody cared.
把模型架构硬编码进芯片是结构性错误；应加速底层的稀疏线性代数。 Hardcoding model architectures into chips is a structural mistake; accelerate underlying sparse linear algebra instead.
最难的问题是封装、供电、散热和数据中心可用性——而非芯片逻辑本身。 Hardest problems are packaging, power, cooling, and data center availability—not the chip logic itself.
推理模型大幅增加推理算力消耗，使速度成为更大的优势。 Reasoning models dramatically increase inference compute, turning speed into an even bigger advantage.
在云中从GPU切换到Cerebras只需八次击键；推理没有CUDA护城河。 Switching from GPU to Cerebras in the cloud takes eight keystrokes; inference has no CUDA moat.
本期章节 · Chapters（共 39）
预告：最大芯片Teaser: The Largest Chip
速度与Token生产力Speed and Token Productivity
Netflix类比The Netflix Analogy
等待的体验The UX of Waiting
AI芯片格局The Landscape of AI Chips
Groq收购与GPU局限Groq acquisition and GPU limitations
OpenAI与芯片需求OpenAI and chip demand
多硅片生态系统Multi-silicon ecosystem
中国在AI领域China in AI
本地AI与芯片Local AI and chips
市场波动Market volatility
需求与泡沫担忧Demand and bubble concerns
存储、封装与晶圆厂瓶颈Memory, packaging, and fab bottlenecks
CPU短缺与代理型AICPU shortage and agentic AI
IPO时机与历程IPO timing and journey
2016年的愿景The Vision in 2016
过往创业的教训Lessons from Previous Ventures
沙漠岁月与晶圆级策略The Desert Years and Wafer-Scale Strategy
真正挑战：封装The real challenge: packaging
里程碑：首颗芯片与IPOMilestones: first working chip and IPO
更大芯片的权衡Tradeoffs of a bigger chip
为何比GPU快Why faster than GPU
解码与硬件优势Decode and Hardware Advantage
快速推理与训练市场Market for Fast Inference and Training
推理与验证Reasoning and Verification
智能体未来Future of Agents
模型验证与多模态Model Verification and Multimodality
服务器业务：芯片、云与数据中心Server Business: Chips, Cloud, and Data Centers
OpenAI数据中心协议与产能限制OpenAI Data Center Deal and Capacity Constraints
数据中心位置、电力与延迟Data Center Location, Power, and Latency
AWS合作与分解式推理AWS Partnership and Disaggregated Inference