OCO-GAT: An Accelerator for Graph Attention Network with Optimized Calculation Order

The Graph Attention Network (GAT) introduces an attention mechanism to focus on the most significant aspects of the data and utilizes weighted sums for aggregation. As a result, GAT exhibits superior performance in tasks involving graph data compared to previous Graph Neural Networks (GNNs). However, this also introduces a more complex computational process and stronger data dependencies, making previous GNN accelerators inadequate for GAT inference. Therefore, we propose an optimized calculation order for GAT along with the corresponding accelerator architecture, OCO-GAT, specifically tailored for GAT inference, which includes efficient pipeline design. Additionally, we introduce a distributed fine-grained on-chip storage scheme, ensuring computational parallelism while mitigating significant growth in storage resource consumption. We deployed OCO-GAT on Xilinx Alveo U250 FPGA and the experimental results demonstrate that optimized calculation order reduces the workload of division operations by an average of 17.2% across four datasets. OCO-GAT achieves high energy efficiency and improves inference performance from 1.2 × to 301.1× compared to CPU, GPU, and three peer works.