ROMJIST Volume 23, No. 3, 2020, pp. 262-273
Voichita DRAGOMIR and Gheorghe M. STEFAN Sparse Matrix-Vector Multiplicationona Map-Reduce Many-Core Accelerator
ABSTRACT: Our proposal for accelerating the computation of Sparse Matrix Vector Multiplication is a Map-Reduce Accelerator as part of a heterogenous computer. We prove that both, structured and unstructured sparse matrices are efficiently multiplied with a dense vector approach using a parallel accelerator structured as a linear array of cells loop connected, through a log-depth reduction network, with a controller. The specific algorithms are presented and their implementation is compared with the of-the-shelf solutions. The main advantages of our architectural proposal, compared with the GeForce GTX 280 GPU which is implemented in the same technological node, are: (1) it provides the means to use 5÷12× more computation out of the peak computational power, (2) it performs the computation with 2.5× less energy.KEYWORDS: sparse matrix, matrix-vector multiplication, unstructured sparse matrix, structured sparse matrix, heterogenous computing.Read full text (pdf)