Communication Kernels: Device-Initiated Transfer, Data Movement, and Compute Fusion

2.14.1GPU P2P Copy, Remote Memory Access, and Memory Registration#

2.14.2Reduction, AllReduce, and ReduceScatter Kernels#

2.14.3NVSHMEM Symmetric Address Space and Explicit Synchronization#

2.14.4Device-Initiated Communication and Host-Proxy Paths#

2.14.5SM-Driven versus DMA/Copy-Engine Data Movement#

2.14.6Communication Tiles, GEMM–Collective Fusion, and Pipelining#

2.14.7SM Occupancy, HBM Contention, and Overlap Limits of Communication Kernels#