NCCL
The collective-communications library every distributed training run calls.
You're looking at a software record: a tool or platform in the AI stack — who maintains it and what hardware it runs on.
NCCL implements the all-reduce, all-gather and broadcast operations that synchronize gradients across thousands of GPUs, tuned to NVLink and InfiniBand topologies. When people say a cluster 'scales well', they are mostly praising NCCL; AMD's RCCL mirrors it.
How this connects to the rest of the stack
This record's links at a glance — click a group to jump to it.
Maintained by
Related
Sources
Record reviewed as of 2026-01-15 — see Methodology.