Unverified50% confidenceFactExact time
业界顶级集群的MFU（Model FLOPS Utilization）通常在35%-60%之间，主要取决于通信拓扑设计、容错调度和系统软件栈的成熟度
1
Sources
50%
Confidence
Long-term
Relevance
7/21/2026
First Seen
Sources
Related Claims
Unverified业界顶级集群的MFU（模型算力利用率）通常在35%-60%之间77% similarUnverified研究表明，即使在 GSM8K 或 MATH 数据集上接近满分的模型，面对稍作变形或需要跨领域组合的问题时性能下降可达30-50个百分点73% similarUnverified工业SCADA故障事件数据高度不平衡，故障事件通常占比不足1%，普通分类模型会偏向正常状态导致低召回率70% similarUnverified混合架构中，仅当规则匹配失败或置信度低于阈值时才调用大模型兜底，通常可减少60%-80%的API调用量69% similarUnverifiedResearch shows that the same base LLM model can vary by 30-50 percentage points in performance under different system prompts.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/575713API
curl https://kongchang.com/api/v1/knowledge/claims/575713MCP
get_claim(id=575713)