Why use NVIDIA GPUDirect Storage in an AI cluster?

NVIDIA GPUDirect Storage is used to create a direct data path between storage and GPU memory. NVIDIA’s GPUDirect Storage Overview Guide states: “GPUDirect® Storage (GDS) enables a direct data path for direct memory access (DMA) transfers between GPU memory and storage, which avoids a bounce buffer through the CPU.” It also says this direct path can relieve system bandwidth bottlenecks and reduce CPU latency and utilization load.

NVIDIA’s GPUDirect Storage technical blog further explains that GPUDirect Storage enables “a direct data path between local or remote storage, like NVMe or NVMe over Fabric (NVMe-oF), and GPU memory.” Therefore, the correct answer is C: it enables peer-to-peer memory transfers between GPUs and NVMe storage.

Why the other options are incorrect: GPUDirect Storage is not primarily about TCP/IP transmission between GPUs and CPUs. It does simplify and bypass parts of the traditional CPU-centric storage path, but the more precise answer is direct GPU-memory-to-storage transfer. It does not increase GPU clock rates.