← The AI Systems Map
L2 · ChipsStandard / protocol⬦ Chokepoint

NCCL / RCCL collectives

Collective-communication library that makes multi-GPU training work

NCCL is NVIDIA's collective-communications library (all-reduce, all-gather) that every large training run depends on to synchronize gradients across thousands of GPUs; AMD's RCCL clones its API. It is an underappreciated part of the CUDA moat because distributed-training performance and fault tolerance hinge on it, and it is tightly co-designed with NVLink and InfiniBand.

Ops

all-reduce, all-gather

AMD equivalent

RCCL

How it fits the stack

NCCL / RCCL collectives with what it depends on (above) and what it feeds (below). The figure renders as a crawlable diagram and upgrades to an interactive 3D graph as it scrolls into view.

depends ondepends ondepends onNCCL / RCCLcollectivesChipsInfiniBand (NVIDIAQuantum)chokepointNVLink / NVSwitchchokepointNvidia Data-Center GPU(Blackwell/Rubin)chokepoint
NCCL / RCCL collectivesDepends on ↑Feeds ↓

NCCL / RCCL collectives in the AI stack. NCCL / RCCL collectives with its immediate upstream dependencies (top) and downstream dependents (bottom) in the AI value chain. Hover a node in 3D, or read the full relationships below.

Graph data (text) — 4 entities, 3 relationships
  • NCCL / RCCL collectivesdepends onInfiniBand (NVIDIA Quantum)
  • NCCL / RCCL collectivesdepends onNVLink / NVSwitch
  • Nvidia Data-Center GPU (Blackwell/Rubin)depends onNCCL / RCCL collectives