Unverified50% confidenceBenchmarkExact time
When actually training Transformer models, the Model FLOPS Utilization (MFU) is typically only 30%–60%.
1
Sources
50%
Confidence
Long-term
Relevance
9/7/2026
First Seen
Sources
Related Entities
Related Claims
Unverified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%69% similarUnverified如果模型跑BF16训练但DCGM_FI_PROF_PIPE_TENSOR_ACTIVE指标长期低于30%，基本可以断定算力被严重浪费66% similarUnverifiedIn the standard Transformer architecture, the feed-forward network (FFN) accounts for approximately two-thirds of the model's total parameter count.66% similarUnverified消费级PWM舵机的连续工作扭矩往往仅为峰值扭矩的50%~60%63% similarUnverified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62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/869484API
curl https://kongchang.com/api/v1/knowledge/claims/869484MCP
get_claim(id=869484)