Neuron Collective Communication · inference-research

Fields

id

sources/alt-neuron-collectives

URL

https://awsdocs-neuron.readthedocs-hosted.com/en/latest/neuron-runtime/about/collectives.html

Publisher

Amazon Web Services

Kind

spec-sheet

Published

2026-08-17

Accessed

2026-10-04

notes

AllGather, ReduceScatter, AllReduce, AlltoAll, Permute. Intra-node (NeuronLink) vs inter-node (EFA). Trn2: 16 chips, 2D torus, 1024 GB/s/chip NeuronLink-v3, 3.2 Tbps EFA. LNC=2 default (4 ranks/chip). CC cores for collective orchestration. 15us RDMA latency.

Notes

AllGather, ReduceScatter, AllReduce, AlltoAll, Permute. Intra-node (NeuronLink) vs inter-node (EFA). Trn2: 16 chips, 2D torus, 1024 GB/s/chip NeuronLink-v3, 3.2 Tbps EFA. LNC=2 default (4 ranks/chip). CC cores for collective orchestration. 15us RDMA latency.

Cross-references

This record references no other records.