Ascend-SpMM: Accelerating Sparse Matrix-Dense Matrix Multiplication on Ascend NPUs

DOI10.48550/ARXIV.2611.01936MaRDI QIDQ7414962arXivFDO

Authors Jiahe Wang, Yingxia Shao, Qingbo Geng, Hongzheng Li, Yucheng Zhang, Shuhan Wang, Qingbo Ge

Publication date 1 November 2026

Published in arXiv

Tldr We present a novel SpMM acceleration framework for modern NPUs featuring a decoupled architecture, introducing a memory-centric cost model and customized sub-kernels with hierarchy-aware pipelining and round-robin synchronization to achieve up to 20.55X speedup over the strongest baseline NPU implementation and competitive GPU performance.

Full work available at URL https://doi.org/10.48550/ARXIV.2611.01936

arXiv classification

cs.DBcs.DC

Cites work

• ASCEND
• Semi-Supervised Classification with Graph Convolutional Networks
• Sparse Matrices in MATLAB: Design and Implementation
• The University of Florida sparse matrix collection
• Two Fast Algorithms for Sparse Matrices: Multiplication and Permuted Transposition

This page was built for publication: Ascend-SpMM: Accelerating Sparse Matrix-Dense Matrix Multiplication on Ascend NPUs

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7414962)