Sense: Model Hardware Co-design for Accelerating Sparse Neural Networks

Sparsity is an intrinsic property of neural network(NN). Many software researchers have attempted to improve sparsity through pruning, for reduction on weight storage and computation workload, while hardware architects are working on how to skip redundant computations for higher energy efciency, but there always exists overhead, causing many architectures suffering from only minor proft. Therefrom, systolic array becomes a promising candidate for the advantages of low fanout and high throughput. However, sparsity is irregular, making it tricky to ft in with the rigid systolic tempo. Thus, this paper proposed a systolic-based architecture, called Sense, for both sparse input feature map(IFM) and weight processing, achieving large performance improvement with relatively small resource and power consumption. Meanwhile, we applied channel rearrangement to gather IFMs with approximate sparsity and co-designed an adaptive weight training method to keep the sparsity ratio(zero element percentage) of each kernel at 1/2, with little accuracy loss. This treatment can effectively reduce the irregularity of sparsity and help better ft with systolic dataflow. Additionally, a dataflow called Partition Reuse is mapped to our architecture, enhancing data reuse, lowering 1.9x-2.6x DRAM access reduction compared with Eyeriss and further reducing system energy consumption. The whole design is implemented on ZynqZCU102 and performs at a peak throughput of 409.6 GOP/s, with power consumption of 11.2W; compared with previous sparse NN accelerators based on FPGA, Sense takes up 1/5 less LUTs and 3/4 less BRAMs, reaches 2.1x peak energy efciency and achieves 1.15x-1.49x speedup.