NVIDIA Optimized primitives for inter-GPU communication (development)

 NCCL (pronounced "Nickel") is a stand-alone library of standard communication
 routines for GPUs, implementing all-reduce, all-gather, reduce, broadcast,
 reduce-scatter, as well as any send/receive based communication pattern. It
 has been optimized to achieve high bandwidth on platforms using PCIe, NVLink,
 NVswitch, as well as networking using InfiniBand Verbs or TCP/IP sockets. NCCL
 supports an arbitrary number of GPUs installed in a single node or across
 multiple nodes, and can be used in either single- or multi-process (e.g., MPI)
 applications.
 .
 This package contains the development files.