<b>Electra:</b> <b>El</b>iminating the In<b>e</b>ffectual <b>C</b>omputations on Bi<b>t</b>map Comp<b>r</b>essed M<b>a</b>trices

The primary computations in several applications, such as deep learning recommendation models, graph neural networks, and scientific computing, involve sparse matrix sparse matrix multiplications (SpMSpM). Unlike standard multiplications, SpMSpMs introduce ineffective computations that can negatively impact performance. While several accelerators have been proposed to execute SpMSpM more efficiently, they often incur additional overhead in identifying the effectual arithmetic computations. To solve this issue, we propose Electra, a novel approach designed to reduce ineffectual computations in bitmap-compressed matrices. Electra achieves this by i) performing logical operations on the bitmap data to know whether the arithmetic computation has a zero or non-zero value, and ii) implementing finer granular scheduling of non-zero elements to arithmetic units. Our evaluations suggest that on average, Electra achieves a speedup of 1.27× over the state-of-the-art SpMSpM accelerator with a small area overhead of 64.92 <inline-formula><tex-math notation="LaTeX">$\text{mm}^{2}$</tex-math></inline-formula> based on 45 nm process.