Unverified50% confidenceFactExact time
Prefill阶段是计算密集型操作，Decode阶段是显存带宽密集型操作
1
Sources
50%
Confidence
Long-term
Relevance
7/12/2026
First Seen
Sources
vLLM推理框架详解：吞吐优化核心原理与面试攻略
bilibiliAI大模型升升6/9/2026
Related Claims
UnverifiedDecode阶段属于访存密集型操作，整体吞吐受制于显存带宽而非算力83% similarUnverified输出Token定价高于输入Token是因为Prefill阶段可并行矩阵运算而Decode阶段必须串行自回归解码，Decode阶段GPU吞吐量约为Prefill阶段的1/6到1/1075% similarUnverifiedDisaggregated Inference架构将Prefill和Decode两个阶段分配到不同的硬件实例上执行74% similarVerified大语言模型推理分预填充和解码两阶段：预填充是计算密集型对GPU算力敏感，解码是内存带宽受限任务71% similarUnverifiedPD分离将推理拆分为Prefill（计算密集、依赖内存容量）和Decode（串行、依赖内存带宽）两阶段，最早由学术界提出，后被字节跳动、微软落地70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/489344API
curl https://kongchang.com/api/v1/knowledge/claims/489344MCP
get_claim(id=489344)