Verified65% confidenceFactTime unknown
自回归解码阶段GPU算力利用率往往不足10%，属于内存密集型操作
3
Sources
65%
Confidence
Long-term
Relevance
6/1/2026
First Seen
Sources
Qwen3.6 MTP加速实测：单GPU推理飙到220 token/s
bilibili普通鱼学家呀
Related Entities
Related Claims
Verified大模型推理时实际GPU利用率往往不足10%85% similarVerified输出Token价格通常为输入的4-6倍，源于预填充与解码两阶段的算力差异，解码阶段GPU利用率可能仅为预填充的10-20%77% similarUnverified纯CPU推理通常比GPU推理慢5-10倍76% similarUnverified传统LLM推理框架为每条请求预分配连续显存导致内存碎片，GPU实际利用率往往不足50%74% similarUnverifiedFreeToken支持运行时调整GPU专家缓存大小，测试显示缓存从58%降至40%仅损失约10%性能，低于20%时性能急剧下降74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/18051API
curl https://kongchang.com/api/v1/knowledge/claims/18051MCP
get_claim(id=18051)