Unverified50% confidenceFactExact time
由N块GPU组成的系统在同一时间段内至少发生一次故障的概率约为1-(1-p)^N，其中p为单块GPU故障概率
1
Sources
50%
Confidence
Long-term
Relevance
7/7/2026
First Seen
Sources
非均匀张量并行：NVIDIA提升LLM训练Goodput的容错新方案
rss7/6/2026
Related Claims
Unverified超大规模分布式训练中硬件故障概率随节点数量指数级上升，GPU 静默错误可能在数百步后才被发现，需从最近检查点重新训练63% similarUnverifiedNon-determinism refers to the phenomenon where the same input may produce slightly different outputs across different runs, which is especially common in models that use GPU parallel computation, dropout layers, or stochastic sampling strategies.63% similarUnverifiedGPTQ通过对每一层权重进行二阶近似优化（基于OBS/OBQ框架）来最小化量化误差，通常需要如C4数据集的128条样本校准，单GPU上仅需数小时完成63% similarUnverifiedFMEA通过失效模式、失效概率及失效后果严重程度三者相乘得出风险优先数（RPN）62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/197026API
curl https://kongchang.com/api/v1/knowledge/claims/197026MCP
get_claim(id=197026)