Race Ya! An Accelerator-Native I/O Pipeline for HDF5
Year: 2025
Type: article
Abstract: GPUs and similar accelerators have become the dominant compute platform for STEM applications, from finance to space flight, and beyond. However, HDF5 continues to execute exclusively on the Host CPU of GPU nodes. This talk will present a design overview of moving the I/O pipeline filters, datatype conversions and other transforms from the CPU to the GPU, including how to perform I/O directly from GPU memory. Implementation details of the revised transform pipeline will be included, along with projected performance benefits.
Author: Koziol, Quincey
Institution: Nvidia (United States)
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6949858910
FWCI:
Cites:
Cited by:
Related to: