H-GAT: A Hardware-Efficient Accelerator for Graph Attention Networks

Recently, Graph Attention Networks (GATs) have shown good performance for representation learning on graphs. Furthermore, GAT leverage the masked self-attention mechanism to get a more advanced feature representation than the graph convolution networks (GCNs). However, GAT incurs large amounts of irregularity in computation and memory access, which prevents the efficient use of traditional neural network accelerators. Moreover, existing dedicated GAT accelerators demand high memory volumes and are difficult to implement onto resource-limited edge devices. Due to this, this paper proposes an FPGA-based accelerator, called H-GAT, which achieves excellent performance on acceleration and energy efficiency in GAT inference. HGAT decomposes GAT operation into matrix multiplication and activation function unit. We first design an effective and fully-pipelined PE for sparse matrix multiplication (SpMM) and dense matrix-vector multiplication (DMVM). Moreover, we optimize the softmax data flow so that the computational efficiency of softmax can be improved dramatically. We evaluate our design on Xilinx Kintex-7 FPGA with three popular datasets. Compared to existing CPU, GPU, and state-of-the-art FPGA-based GAT accelerator, H-GAT can achieve speedup by up to 585×, 2.7×, and 11× and increases power efficiency by up to 2095×, 173×, and 65×, respectively.