Μελέτη απόδοσης συστήματος PIM (Process in Memory) ως προς την υλοποίηση πολλαπλασιασμού αραιών μητρώων επί διάνυσμα SPMV
Study of PIM (Process in Memory) system performance regarding Sparse Matrix-Vector Multiplication SPMV implementation
View/
Abstract
Computational performance is a frequent concern when it comes to sparse matrix-vector multiplication (SPMV). In such multiplications, the execution time increases significantly. The problem is intensified in sparse matrices that include 64-bit double-precision floating-point elements, which are decimal numbers with many digits before or after the decimal point. The architecture and, consequently, the mode of operation of central processing units (CPUs) are not sufficient to perform the operations in a reasonable amount of time. In contrast, graphics processing units (GPUs), with their ability to perform parallel computations, are now the main solution to the problem as they execute these operations more efficiently.
The purpose of this work is to study a new processor architecture, Process In Memory (PIM), which not only performs parallel computations but also differs in that it integrates a processing unit and memory into a single circuit. This avoids the delays that arise from the transfer of information between the two.
For this purpose, the performance of the PIM system was compared with that of a GPU. Specifically, the execution time was measured for the two alternatives (GPU and PIM) for sparse matrix-vector multiplication on thirteen different matrices using two alternative operating techniques for the PIM system. The above was applied to a real scientific application that studies the simulation of biological neuron networks (Leaky Integrated-and-Fire (LIF)).
This work suggests that this technology needs improvements to become usable for real scientific applications.