Abstract
I/O is a bottleneck in GPU-accelerated science, yet scientific data-format libraries run only on the host. A kernel cannot write where its data lives: it must exit and copy its array before HDF5 can run, so the write point becomes a kernel boundary. Systems that let a GPU initiate I/O give up the data model, emitting flat, untyped byte ranges. We present GPUH5, a GPU-first I/O library for persistent kernels. It keeps HDF5's hierarchical paths and chunked arrays, replacing the container format with a key-value store, so the host precomputes every chunk's key and an in-kernel write reduces to enqueuing a pointer, light enough that the producer holds 98.8% achieved occupancy. On a single-GPU Gray-Scott simulation writing 380 snapshots to NVMe, GPUH5 hides 96% of the I/O its synchronous twin exposes, finishing 1.9× faster than default HDF5.