A GPU file read can use the right API and still stage through CPU memory

NVIDIA’s KvikIO tutorial shows that reading directly into a GPU buffer does not prove the transfer bypassed CPU memory. KvikIO first chooses between a POSIX path and cuFile; cuFile then independently chooses accelerated GPUDirect Storage or a compatibility path. The walkthrough uses per-process cuFile counters to identify what actually happened. This turns a performance assumption into an observable property of the storage and GPU configuration.

NVIDIA Technical Blog · Oct 22, 2026

Why it matters

Layered fallback paths can preserve correctness while hiding the acceleration an application expected to use.

More context

The example requires Linux, an NVIDIA GPU, GPUDirect Storage and a supported storage/filesystem configuration. NVIDIA tested an H100 with CUDA 13.3, cuFile 1.18 and local ext4 NVMe storage. Try the KvikIO example on an already compatible system and compare nvfs with posix counters.

Keep in mind

Topology, request size and both libraries’ settings affect the selected path. The tutorial demonstrates path verification rather than a universal throughput gain; merely forcing KvikIO to call cuFile is insufficient.

Original read