For ML training collectives, a Clos network usually beats a torus except for AllGather, where the torus's higher per-node access bandwidth wins.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.NI 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
On Topology's Role in ML Training Performance
For ML training collectives, a Clos network usually beats a torus except for AllGather, where the torus's higher per-node access bandwidth wins.