Unverified50% confidenceFactExact time
nvidia-smi报告的GPU利用率本质上只统计采样时间窗口内GPU上至少有一个内核在执行的时间占比，而非计算资源的实际占用率
1
Sources
50%
Confidence
Long-term
Relevance
8/7/2026
First Seen
Sources
Related Claims
UnverifiedDuring testing, only 1 of the 4 GPUs reached a utilization rate of approximately 10%–13%, while the rest were nearly idle, and CPU usage was as high as approximately 50%.73% similarUnverified单条推理通常只能占用GPU SM利用率的5%～15%，而合理的批次大小可将利用率提升至60%～90%73% similarUnverified对于MuJoCo、Atari等环境，单个环境的采样速度往往跟不上GPU计算能力，导致GPU利用率长期停留在20%–40%甚至更低71% similarUnverifiedDuring the autoregressive decoding phase, GPU compute utilization is often less than 10%, making it a memory-intensive operation.71% similarUnverified相较于用GPU执行AI推理，NPU的功耗通常只有前者的1/5到1/1071% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/700094API
curl https://kongchang.com/api/v1/knowledge/claims/700094MCP
get_claim(id=700094)