The issue of financial fraud is attracting increasing social attention. The task of Graph Fraud Detection (GFD) is a typical application of Graph Neural Networks (GNNs) in the e-commerce field. Although several successful works have already achieve excellent performance, there are still many obstacles before these methods can be used in financial market industry. Many of the GFD methods are based on spatial approaches, such as neighbor selection, which will slow down their execution on large datasets. This poses problems in e-commerce settings where real-time detection is essential. Also, fraudsters often disguise their behavior, resulting in high heterophily of the graph, which makes direct aggregation of neighbors less effective. To address these issues, we introduce the Fast Cluster Multi-Hop Detector (FCMH). This model first partitions the graph into different small subgraphs to accelerate the neighbor aggregation process. Then, it uses a multi-hop neighborhood aggregation strategy to simultaneously aggregate multiple layers of neighbors to learn fraudulent patterns. Our model has been evaluated on both the open source and industrial e-commerce datasets, and it has produced ideal results in terms of classification performance and time efficiency. We believe that our work will be beneficial for the applications of GFD models in financial fraud detection.