Unverified50% confidenceFactExact time
大语言模型推理中输出 token 的实际硬件成本远高于输入 token，因解码阶段是内存带宽受限的串行过程
1
Sources
50%
Confidence
Long-term
Relevance
7/9/2026
First Seen
Sources
GPT-5.6 Sol限量预览：分级模型Sol/Tera/Luna与准入机制详解
bilibili开水庆余年6/27/2026
Related Claims
Unverified语言模型生成 token 本质上是内存带宽受限任务而非纯算力受限81% similarUnverifiedThe computational cost of generating output is typically far higher than that of processing input, and some vendors set separate quotas for input tokens and output tokens.74% similarUnverifiedTokens are the basis for billing by large models; they are produced by a tokenizer splitting raw text and are the basic unit by which large models process text.73% similarVerified大语言模型的推理过程本质上是内存带宽受限（memory-bandwidth bound）的任务，每生成一个Token都需要将模型的全部权重从存储中读取一遍72% similarUnverifiedIn coding scenarios, the input side needs to carry a large amount of code context, and the input Token count is often several times that of the output, amplifying costs.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/366702API
curl https://kongchang.com/api/v1/knowledge/claims/366702MCP
get_claim(id=366702)