NCCL, the NVIDIA Collective Communications Library, implements optimized collective operations such as all-reduce, all-gather, and broadcast across many GPUs. It automatically chooses efficient communication paths over NVLink, NVSwitch, and InfiniBand, and underpins the gradient synchronization in distributed training frameworks. NCCL performance is often the deciding factor in how well a GPU cluster scales, making it central to interconnect and topology design.
Category: IT & Compute Infrastructure · Subcategory: AI & HPC Infrastructure · Part of the AIDC Technical Terminology reference.