Details
- Affiliation
-
AffiliationImperial College London
- Country
Graph Convolutional Networks (GCNs) have become a promising approach in the graph data processing and analysis. Nevertheless, the massive amount of input graph nodes and computation put a heavy burden on their hardware performance, limiting their their deployment in real-life applications. To address this performance issue, we propose a customizable architecture for Binarized GCNs (BiGCNs), which can be reconfigured to fulfill different user needs with high hardware performance. We overlap the binarization and computation to ease the bandwidth requirement for BiGCNs. To further improve the hardware performance, we also adopt COO(Coordinate) format storage for the adjacency matrix to skip the redundant computation for higher hardware performance. Our FPGA design,for the Cora and Flickr datasets,can run 77 and 16 times faster than the corresponding CPU and GPU platforms.