NVIDIA published a tutorial on reading data into GPU memory from Python using KvikIO, the Python and C++ library that sits above the cuFile API for GPUDirect Storage. The post describes GDS as a direct data path between supported storage and GPU memory that avoids staging through CPU memory and reduces CPU utilisation, and notes that KvikIO can fall back to a POSIX compatibility path that stages data through host memory. It also states that reaching the cuFile API does not by itself confirm the accelerated path was used, because cuFile decides the transfer route from system support, file system and configuration. The tutorial walks through configuring storage I/O, issuing reads into a GPU buffer and inspecting the KvikIO and cuFile settings that determine which path an operation takes. As context, the post cites the 2025 Stanford AI Index for compute for notable models doubling roughly every five months and LLM training dataset size doubling roughly every eight months, and says nearly half of new AI projects on GitHub were built primarily in Python as of August 2025.
Our takeThe useful part is verification, not the API call. A program can appear to use GPUDirect while landing on a POSIX path or on cuFile's CPU-staged compatibility route, so logging which path each read actually took belongs in the benchmark, not the appendix. The cited trend numbers, compute doubling about every five months and training data every eight, are aggregate AI Index figures rather than a claim about any single pipeline's I/O demand. Worth noting that the speed-up depends on the file system, driver and storage stack, none of which the tutorial generalises over.