Unverified50% confidenceFactExact time
批大小为1的Transformer推理算术强度往往低于10 FLOP/Byte，意味着GPU算力利用率不足理论峰值的5%
1
Sources
50%
Confidence
Long-term
Relevance
7/8/2026
First Seen
Sources
Related Claims
UnverifiedTransformer推理中批量大小为1时注意力层的算术强度约1~10 FLOP/Byte，远低于H100约300 FLOP/Byte的屋脊点，属于带宽受限场景79% similarUnverified将FP32权重降至INT8后，典型Transformer模型基准测试分数下降通常不超过1%，推理速度提升2至4倍，内存占用减少75%74% similarUnverifiedDuring the autoregressive decoding phase, GPU compute utilization is often less than 10%, making it a memory-intensive operation.71% similarUnverifiedDuring large model inference, actual GPU utilization is often less than 10%.69% 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1/1067% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/229424API
curl https://kongchang.com/api/v1/knowledge/claims/229424MCP
get_claim(id=229424)