What is it about?
We propose MatXtract to optimize Sparse Matrix-Vector multiplication (SpMV) through sparsity-aware matrix reordering and dense computation extraction. Instead of a "one-size-fits-all" approach, MatXtract leverages Bayesian optimization and tree models to analyze specific sparse features. It adaptively predicts the suitable reordering strategy and distributes computing tasks between Tensor Cores and CUDA Cores. In evaluations on an NVIDIA A100 GPU covering 2,059 real-world sparse matrices, MatXtract outperforms cuSPARSE in 96.64% of cases, achieving an average speedup of 1.98× and up to 8.83×.
Featured Image
Photo by Growtika on Unsplash
Why is it important?
SpMV is a critical operator for workloads like AI for Science, graph analytics, and sparse linear solvers. However, a major challenge exists: modern GPUs are optimized for dense computation, while real-world data is often highly irregular and sparse. This mismatch typically causes low utilization of cache and computing resources. MatXtract mitigates this and offers a practical, high-performance SpMV system to accelerate sparse workloads across various sparse patterns.
Perspectives
Sparsity should be treated as an opportunity for optimization, not merely a performance bottleneck.
Luhan Wang
Peking University
The real challenge is not sparsity itself, but how to uncover the computational value hidden within it.
Kun Li
Tsinghua University
Read the Original
This page is a summary of: MatXtract: Sparsity-Aware Matrix Transformation via Cascaded Compute Density EXtraction for SpMV, ACM Transactions on Architecture and Code Optimization, January 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3793864.
You can read the full text:
Contributors
The following have contributed to this page
