Flexagon: A Multi-Dataflow Sparse-Sparse Matrix MultiplicationAccelerator for Efficient DNN Processing

Abstract

Sparsity is a growing trend in modern DNN models.Existing Sparse-Sparse Matrix Multiplication (SpMSpM) accel-erators are tailored to a particular SpMSpM dataflow (i.e., InnerProduct, Outer Product or Gustavson’s), which determines theiroverall efficiency. We demonstrate that this static decision inher-ently results in a suboptimal dynamic solution. This is becausedifferent SpMSpM kernels show varying features (i.e., dimensions,sparsity pattern, sparsity degree), which makes each dataflow bettersuited to different data sets.In this work we present Flexagon, the first SpMSpM reconfig-urable accelerator that is capable of performing SpMSpM computa-tion by using the particular dataflow that best matches each case.Flexagon accelerator is based on a novel Merger-Reduction Net-work (MRN) that unifies the concept of reducing and merging inthe same substrate, increasing efficiency. Additionally, Flexagonalso includes a new L1 on-chip memory organization, specificallytailored to the different access characteristics of the input and out-put compressed matrices. Using detailed cycle-level simulation ofcontemporary DNN models from a variety of application domains,we show that Flexagon achieves average performance benefits of4.59×, 1.71×, and 1.35×with respect to the state-of-the-art SIGMA-like, SpArch-like and GAMMA-like accelerators (265%, 67%, and18%, respectively, in terms of average performance/area efficiency).