Unverified50% confidenceFactExact time
The Decode stage produces only one token at a time, resembling GEMV in form, with extremely low arithmetic intensity and is memory-bound.
1
Sources
50%
Confidence
Long-term
Relevance
9/23/2026
First Seen
Sources
Related Entities
Related Claims
Unverified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1/1068% similarUnverifiedGPU utilization during the decoding phase of large language model inference is far lower than during the prefill phase's parallel matrix computation, because decoding is a step-by-step autoregressive generation process64% similarUnverifiedAt the hardware level, the Prefill (input) stage can read all tokens in parallel and compute the self-attention matrix simultaneously, whereas the Decode (output) stage requires serial autoregressive token-by-token inference, causing GPU utilization to drop sharply.64% similarUnverified自回归解码具有串行特性，每个Token的生成依赖前一个Token，难以并行64% similarUnverified生成式AI推理中的自回归解码本质上是内存带宽受限（Memory-Bound）的操作，每次生成Token需读取全部KV Cache64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/942615API
curl https://kongchang.com/api/v1/knowledge/claims/942615MCP
get_claim(id=942615)