Chinese research team develops breakthrough solver for sparse matrix efficiency using Kunpeng processors
A research team led by Chen Junshi at the University of Science and Technology of China has developed a novel LU direct solver that dramatically improves the efficiency of sparse matrix computations. The breakthrough was achieved using the matrix computing power of Huawei's Kunpeng processors, specifically the Kunpeng 920 model.
Sparse matrix equations are central to scientific computing, but traditional solvers often suffer from inefficiencies due to the irregular distribution of non-zero elements. These inefficiencies lead to wasted computing power and slow performance.
To overcome this, the team designed a dense-kernel LU solver tailored for Kunpeng's architecture. Their approach transforms irregular sparse matrices into unified dense block structures, allowing for more streamlined numerical decomposition. This method replaces fragmented operations with dense matrix routines, significantly reducing memory access overhead and improving scheduling.
The Kunpeng 920 processor plays a critical role in this solution. It features a dedicated matrix computing unit optimized for high-density numerical operations and a multi-core architecture that supports parallel processing of large data blocks. Combined with the Kunpeng KML math library, which is tuned for dense matrix workloads, the system achieves high performance through tight hardware-software integration.
In benchmark tests across more than 70 datasets, the new solver delivered an average speedup of 9.6x and peak acceleration of 32.2x compared to the widely used SuperLU solver. This significantly reduces the time cost of scientific computing tasks.
The research has been published at Euro-Par 2025, a CCF Class B international conference, and the core technology has already received a Chinese invention patent.
The LU solver was developed by Chen Junshi's team at the University of Science and Technology of China, with hardware support from Huawei's Kunpeng Ascend Science and Education Innovation Center.
