Polyomino: A 3D-SRAM-Centric Architecture for Randomly Pruned Matrix Multiplication with Simple Rearrangement Algorithm and x0.37 Compression Format
Kota Shiba, Mitsuji Okada, Atsutake Kosuge, Mototsugu Hamada, Tadahiro Kuroda · 2022 20th IEEE Interregional NEWCAS Conference (NEWCAS) · 2022
We propose a sparse matrix rearrangement algorithm with a novel 3D-SRAM-centric Polyomino architecture which makes it possible to efficiently process the rearranged matrix for the compression of parameters. By rearranging randomly pruned, irregularly structured sparse matrices into regularly structured matrices, the compression ratio of the data increases, as well as the efficiency of hardware processing. The rearrangement algorithm can be implemented simply by attributing it to the widely known k-sum problem. We also propose a compression format for storing the rearranged matrices and show that the rearranged regular structure can reduce the amount of required memory by 63% compared with the conventional method. The proposed Polyomino architecture can efficiently process rearranged matrices by using a 3D stacked SRAM, which is an external memory with random accessibility and low latency.