GPUDirect Storage and NVMe-oF

GPUDirect Storage (GDS) allows data to move directly between storage devices and GPU memory, bypassing the CPU:

• Eliminates CPU bottleneck in data loading
• Dramatically improves I/O throughput
• Requires: compatible NVIDIA GPU + driver, nvidia_fs kernel module, compatible NVMe/NVMe-oF storage

gdscheck -p           # Check GDS availability
modprobe nvidia_fs    # Load the kernel module if not loaded
lsmod | grep nvidia_fs

NVMe-oF (NVMe over Fabrics) extends NVMe protocol over a network:

• NVMe/RDMA — Over InfiniBand or RoCE (lowest latency)
• NVMe/TCP — Over standard Ethernet (easier to deploy)
• Provides near-local NVMe performance across the network