NCCL

This guide walks through building a full NVIDIA GPU compute node on Ubuntu 20.04/22.04, targeting HGX A800 and Tesla data center GPUs. It covers disabling Nouveau, installing the NVIDIA driver and CUDA, configuring Infiniband with MLNX OFED, enabling key kernel modules, building OpenMPI, setting up NCCL and FabricManager, and validating with single-node and multi-node NCCL+MPI benchmarks.