Unverified50% confidenceFactExact time
Hierarchical networks often adopt a design of high bandwidth within racks and oversubscription between racks, but the global communication pattern of All-Reduce raises the share of cross-layer traffic, thereby amplifying the oversubscription bottleneck.
1
Sources
50%
Confidence
Long-term
Relevance
10/28/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn distributed inference, pipeline parallelism (splitting by layer) has lower network bandwidth requirements but uneven device utilization; tensor parallelism (splitting the matrix operations of a single layer) communicates frequently and is latency-sensitive.78% similarUnverified张量并行每一层前向计算几乎都要执行一次all-reduce等集合通信，通信频率随层数线性增长，对带宽和延迟极度敏感。78% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1109543API
curl https://kongchang.com/api/v1/knowledge/claims/1109543MCP
get_claim(id=1109543)