GPUDirect is a family of NVIDIA technologies that let GPUs exchange data directly with network adapters, storage, and other GPUs without staging copies through host CPU memory. GPUDirect RDMA moves data straight between GPU memory and a NIC, while GPUDirect Storage streams data from NVMe or a parallel file system into GPU memory. In AI clusters it cuts latency and CPU overhead, keeping accelerators fed during training and inference.
Category: IT & Compute Infrastructure · Subcategory: AI & HPC Infrastructure · Part of the AIDC Technical Terminology reference.