Performance Evaluation of SpGEMM on RISC-V Vector Processors at the Barcelona Supercomputing Center

Time Stamp: March 10, 2023 2:12 PM
Source Node: 2016187

The Barcelona Supercomputing Center (BSC) has recently conducted a performance evaluation of the SpGEMM algorithm on RISC-V vector processors. The SpGEMM algorithm is a sparse matrix-matrix multiplication algorithm that is used to speed up the calculation of large matrix operations. The performance evaluation was conducted on the MareNostrum 4 supercomputer, which is equipped with the latest RISC-V vector processors.

The performance evaluation was conducted by running the SpGEMM algorithm on a range of different matrix sizes, ranging from small matrices of size 16x16 to large matrices of size 8192x8192. The results of the evaluation showed that the SpGEMM algorithm was able to achieve a significant speedup over traditional dense matrix-matrix multiplication algorithms. For example, when multiplying two 8192x8192 matrices, the SpGEMM algorithm was able to achieve a speedup of up to 8.5x over traditional algorithms.

The BSC researchers also evaluated the performance of the SpGEMM algorithm on different types of matrices. They found that the algorithm performed best on matrices with a high degree of sparsity, such as those found in scientific computing applications. For example, when multiplying two 8192x8192 matrices with a sparsity of 0.9, the SpGEMM algorithm was able to achieve a speedup of up to 11x over traditional algorithms.

Overall, the performance evaluation conducted by the BSC researchers showed that the SpGEMM algorithm is an effective way to speed up large matrix operations on RISC-V vector processors. The algorithm was able to achieve significant speedups over traditional algorithms, and it performed best on matrices with a high degree of sparsity. This makes it an ideal choice for scientific computing applications that require fast matrix operations.

Time Stamp:

More from Semiconductor / Web3