Unverified50% confidenceFactExact time
In the Transformer, the number of subspace directions drops from more than a quarter of the available directions to about one tenth during the phase transition.
1
Sources
50%
Confidence
Long-term
Relevance
10/27/2026
First Seen
Sources
Related Entities
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65% similarUnverified基于Transformer架构的现代审核模型相比传统关键词过滤，误报率（False Positive Rate）可降低60%以上65% 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%63% similarUnverified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%63% similarUnverifiedConformer achieved approximately 15% relative error rate reduction on the LibriSpeech benchmark compared to a pure Transformer architecture.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1102243API
curl https://kongchang.com/api/v1/knowledge/claims/1102243MCP
get_claim(id=1102243)