Unverified50% confidenceFactExact time
DCGM_FI_PROF_SM_ACTIVE measures the proportion of all SMs that are actually in an active computing state, while DCGM_FI_PROF_PIPE_TENSOR_ACTIVE measures the actual utilization rate of tensor compute units.
1
Sources
50%
Confidence
Long-term
Relevance
9/3/2026
First Seen
Sources
Related Entities
Related Claims
Unverified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62% similarUnverified选择性重计算由Korthikanti等人在2022年的Megatron-LM论文中系统研究，仅对注意力激活值做重计算可在额外计算开销不超过10%的条件下接近全量重计算的显存节省56% similarUnverifiedLLM每次调用可建模为伯努利试验，N次调用通过次数服从二项分布B(N, p)，通过率95%置信区间约为p̂ ± 1.96×√(p̂(1-p̂)/N)56% similarUnverifiedIn a scenario where GPU Util is 100%, SM Active may be only 15% and Tensor Core Active may be less than 5%, with the granularity difference among the three reaching orders of magnitude.55% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/849968API
curl https://kongchang.com/api/v1/knowledge/claims/849968MCP
get_claim(id=849968)