← The AI Systems Map
L2 · ChipsFramework / tooling⬦ ChokepointSanta Clara, CA

NCCL (NVIDIA Collective Comms)

The collective-communications library gluing multi-GPU training

NCCL is NVIDIA's collective-communications library that makes all-reduce/all-gather across thousands of GPUs performant; it is the invisible dependency of every large training run and reinforces the CUDA moat alongside cuDNN. Competitors must reimplement it (RCCL for AMD) to be viable.

Role

Multi-GPU all-reduce

AMD analog

RCCL

How it fits the stack

NCCL (NVIDIA Collective Comms) 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 onNCCL (NVIDIACollective Comms)ChipsCUDA / software moatchokepointNvidia Data-Center GPU(Blackwell/Rubin)chokepoint
NCCL (NVIDIA Collective Comms)Depends on ↑Feeds ↓

NCCL (NVIDIA Collective Comms) in the AI stack. NCCL (NVIDIA Collective Comms) 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) — 3 entities, 2 relationships
  • NCCL (NVIDIA Collective Comms)depends onCUDA / software moat
  • Nvidia Data-Center GPU (Blackwell/Rubin)depends onNCCL (NVIDIA Collective Comms)