SkeletonGCN: A Simple Yet Effective Accelerator For GCN Training

Graph Convolutional Networks (GCNs) have shown great results but come with large computation costs and memory overhead. Recently, sampling-based approaches have been proposed to alter input sizes, which allows large GCN workloads to align to hardware constraints. Motivated by this flexibility, we propose an FPGA-based GCN accelerator, named SkeletonGCN, along with multiple software-hardware co-optimizations to improve training efficiency. We first quantize all feature and adjacency matrices of GCN from FP32 to SINT16. We then simplify the non-linear operations to better fit the FPGA computation, and identify reusable intermediate results to eliminate redundant computation. Moreover, we employ a linear time sparse matrix compression algorithm to further reduce memory bandwidth while allowing efficient decompression on hardware. Finally, we propose a unified hardware architecture to process sparse-dense matrix multiplication (SpMM) and dense matrix multiplication (MM), all on the same group of PEs to increase DSP utilization on FPGA. Evaluation is performed on a Xilinx Alveo U200 board. Compared with existing FPGA-based accelerator on the same network architecture, SkeletonGCN can achieve up to 11.3x speedup while maintaining the same training accuracy. In addition, SkeletonGCN can achieve up to 178x and 13.1x speedup over state-of-art CPU and GPU implementation on popular datasets, respectively.