NVIDIA Introduces DOCA GPUNetIO to Unify GPU Networking

NVIDIA introduced DOCA GPUNetIO, a software development kit that brings GPU-initiated networking into a common framework across NVIDIA’s software ecosystem.

The toolkit allows CUDA kernels to handle Ethernet, Remote Direct Memory Access (RDMA), Verbs and direct memory access operations. That reduces the need for the CPU to manage every network transaction, supporting applications that require high efficiency and low latency in artificial intelligence, data analytics and high-performance computing (HPC).

GPUNetIO consolidates previously siloed GPU-initiated, kernel-initiated networking implementations, known as GDA-KI, across NVIDIA libraries including NCCL, NVSHMEM, UCX/NIXL and Holoscan. A shared framework can reduce duplicated code and make software improvements available across more of the stack.

NVIDIA’s NCCL GIN, or GPU-Initiated Networking, uses GPUNetIO for RDMA operations in collective algorithms. The integration is designed to enhance bandwidth and reduce latency in those workloads.

Developers can use GPUNetIO through an open-source implementation focused on RDMA and Verbs or through the broader DOCA software development kit. The DOCA version adds enhanced RDMA, Ethernet and DMA capabilities.

The two implementations support different levels of functionality. Applications can begin with a narrower networking implementation and expand to more advanced capabilities through the broader DOCA toolkit.