NVIDIA/nccl

Health v2 · maintenance only

Flags: no_license

Adoption not part of the score

5027 stars · 1400 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

NVIDIA's Collective Communication Library (NCCL) is a C++ library providing topology-aware, high-bandwidth inter-GPU communication primitives such as AllReduce, AllGather, Broadcast, and point-to-point send/receive. It is optimized for PCIe, NVLink, NVSwitch, InfiniBand, and TCP/IP fabrics and integrates into single- or multi-process (e.g., MPI) applications.

Use cases

• synchronize gradients across multiple GPUs for deep learning training
• run allreduce collectives across nodes with InfiniBand
• scale model training to multi-GPU and multi-node clusters
• implement custom send/receive communication patterns between GPUs
• accelerate distributed training frameworks like PyTorch or TensorFlow

When to choose

• you need maximum-bandwidth GPU-to-GPU collectives on NVIDIA hardware
• you are building or tuning a distributed training framework
• you need collectives over NVLink, NVSwitch, or InfiniBand fabrics
• you want an MPI-like collective API with CUDA stream semantics

When to avoid

• you need secure or encrypted network communication between GPUs
• you need a full parallel programming framework with process launching
• your workload runs on non-NVIDIA GPUs
• you only need CPU-to-CPU message passing, where MPI alone suffices

Facets

library · maturity stable

concurrency networking gpu-computing deep-learning microservices gpu-computing cpp collective-communication mpi-alternative nvidia multi-gpu nvlink infiniband allreduce hpc high-performance-computing linux cuda gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="NVIDIA/nccl")