Graph Neural Networks (GNNs) are a type of deep neural network that can learn directly from irregular graph-structured data. However, GNN inference relies on sparse operations that are difficult to implement on accelerator architectures optimized for dense, fixed-sized computation. This paper proposes GReTA (Gather, Reduce, Transform, Activate), a graph processing abstraction designed to be efficient for an accelerator to execute and flexible enough to implement GNN inference. We demonstrate GReTA’s advantages by designing and synthesizing a custom accelerator ASIC for GReTA and implementing several GNN models (GCN, GraphSage, G-GCN, and GIN.) Across several benchmark graphs, our implementation reduces 97th percentile latency by a geometric mean of 15× and 21× compared to a CPU and GPU baseline respectively.