Dylan Patel shares his journey from growing up in a family motel and gas station to founding Semi Analysis, a premier research firm in the semiconductor space.
要点 · TL;DR
推理将成为全球最大市场之一，超过石油。 Inference will become one of the largest global markets, surpassing oil.
软硬件协同设计在堆栈各层带来倍增收益。 Hardware-software co-design yields multiplicative gains across the stack.
CUDA 护城河正在减弱，模型公司为多种芯片编写自定义内核。 CUDA moat is weakening as model companies write custom kernels for multiple chips.
核心观点 · Key points
推理将成为全球最大市场之一，超越石油，贡献显著的GDP份额。 Inference will become one of the largest markets globally, surpassing oil and contributing significant GDP share.
硬件-软件协同设计至关重要；跨模型、软件和硬件层的联合优化能带来乘数级收益。 Hardware-software co-design is critical; co-optimizing across model, software, and hardware layers yields multiplicative gains.
内存带宽是关键瓶颈；直接芯片堆叠等创新将大幅提升性能。 Memory bandwidth is a key bottleneck; innovations like direct chip stacking will dramatically improve performance.
新云机会存在，因为超大规模云的传统基础设施对AI工作负载并非最优。 The neocloud opportunity exists due to hyperscalers' legacy infrastructure being suboptimal for AI workloads.
黄仁勋推动多极化世界，支持多样化的AI实验室和新云，以减少对超大规模云的依赖。 Jensen Huang promotes a multipolar world with diverse AI labs and neoclouds to reduce dependency on hyperscalers.
反共识 · Contrarian takes
太空数据中心在未来3-5年并不重要；20年后大部分算力才会转移到太空。 Space data centers will not matter in 3-5 years; only after 20 years will majority of compute move to space.
CUDA护城河正在减弱；模型公司现在为多种芯片编写自定义内核，借助AI编码。 CUDA moat is weakening; model companies now write custom kernels for multiple chips, aided by AI coding.
过去三年AI性能提升主要来自模型层，而非硬件；协同设计是关键。 Most AI performance gains in last 3 years came from model layer, not hardware; co-design is key.
英伟达和TPU无法直接比较；最优选择取决于模型架构的协同设计。 Nvidia and TPU are not directly comparable; optimal choice depends on model architecture co-design.
如果模型进步超过算力增长，算力紧缺可能无限持续，但资本约束可能减缓。 Compute crunch may persist indefinitely if model progress outpaces compute growth, but capital constraints could slow it.
AI有明确回报；否认模型进步是常见误解，尽管能力持续提升。 AI has clear ROI; denial of model progress is a common misconception despite continuous capability improvements.
本期章节 · Chapters（共 21）
引言与半分析文化Introduction and Semi Analysis Culture
早年生活与首个神经网络Early Life and First Neural Network
Xbox 360与硬件兴趣Xbox 360 and Interest in Hardware
痴迷与成绩Obsession and Grades
创办半分析Starting Semi Analysis
会议经历与供应链洞察Conference experiences and supply chain insights
推理基准测试与吞吐-交互曲线Inference benchmarking and the throughput-interactivity curve
每瓦智能与协同优化Intelligence per watt and co-optimization
全栈协同优化Co-optimization across the stack
软硬件协同设计与模型架构分化Hardware-Software Co-Design and Model Architecture Divergence
CUDA护城河与生态变迁CUDA Moat and Changing Ecosystem
对Cerebras与快速推理的思考Thoughts on Cerebras and Fast Inference
Cerebras与大模型推理挑战Cerebras and large model inference challenges
长期押注与遇见DeaveenLong-term bet and meeting Deaveen
数据中心建设与算力紧缺Data center buildout and compute crunch
算力紧缺与模型改进Compute crunch and model improvement
数据中心质量与定价差异Data center quality and pricing disparity
AI实验室的供需约束Demand and supply constraints at AI labs
Neocloud与超大规模云优势Neocloud vs hyperscaler advantages
黄仁勋的多极世界策略Jensen's strategy for a multipolar world