Extending Ginkgo to Manage Reconfigurable Hardware-Based Kernels
Although heterogeneous systems based on hardware accelerators are a trending topic in the HPC community, exploring the trade-offs of reconfigurable hardware-based ones in linear algebra libraries for high-performance systems, has not been deeply studied. Therefore, in this research, we aim to take...
Saved in:
Main Authors: | , |
---|---|
Format: | Article |
Language: | English |
Published: |
Universidad Autónoma de Bucaramanga
2025-01-01
|
Series: | Revista Colombiana de Computación |
Subjects: | |
Online Access: | https://revistas.unab.edu.co/index.php/rcc/article/view/5276 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Summary: | Although heterogeneous systems based on hardware accelerators are a trending topic in the HPC community, exploring the trade-offs of reconfigurable hardware-based ones in linear algebra libraries for high-performance systems, has not been deeply studied. Therefore, in this research, we aim to take advantage of FPGAs' reconfigurability, adaptability, and capacity to reduce power consumption to generate FPGA-based kernels in Ginkgo, a specialized high-performance linear algebra library for many-core systems. We generated 3 FPGA-based kernels for the CSR, SELLP, and SELL SpMV formats, and obtained speedups of at least 10x concerning CPU-based kernels. Furthermore, we demonstrated via a performance characterization study that FPGAs outperform general-purpose processors in terms of compute time.
|
---|---|
ISSN: | 1657-2831 2539-2115 |