Unverified50% confidenceFactExact time
At 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.
1
Sources
50%
Confidence
Long-term
Relevance
9/2/2026
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1/1077% 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 process76% similarVerifiedLarge language model inference is divided into two phases, prefill and decode: prefill is compute-intensive and sensitive to GPU processing power, while decode is a memory-bandwidth-bound task.76% similarUnverifiedThe Prefill phase is a compute-intensive operation, while the Decode phase is a memory-bandwidth-intensive operation.71% similarUnverifiedThe inference process of large language models is divided into two phases, Prefill and Decode, and the computational complexity of the attention mechanism is O(n²).70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/842317API
curl https://kongchang.com/api/v1/knowledge/claims/842317MCP
get_claim(id=842317)