NVIDIA consolidates GPU networking across communication libraries

NVIDIA says GPUNetIO now provides a shared foundation for GPU-initiated networking across libraries including NCCL, NVSHMEM and UCX/NIXL, replacing separate implementations. Its software lets CUDA kernels submit network operations and notify network cards directly, keeping the CPU out of the application's critical data path. The shared implementation lets participating frameworks reuse networking features and fixes. NVIDIA publishes an open-source version focused on RDMA Verbs, which supports direct transfers between device memories; the full DOCA SDK adds broader Ethernet and DMA capabilities. Applications still require an initial CPU configuration phase before GPU kernels handle traffic.