GCoD: Graph Convolutional Network Acceleration via Dedicated Algorithm and Accelerator Co-Design

TL;DR

This work targets the efficiency bottleneck of GCN inference on large graphs by introducing GCoD, an algorithm–accelerator co-design that polarization-izes adjacency into two regular workloads (denser and sparser neighborhoods) and processes them with a dedicated two-branch accelerator. The algorithm comprises a split-and-conquer training strategy with group-class subgraphs, patch-based structural sparsification, and an early-stopping pipeline to minimize training overhead, while the hardware design features a Denser Branch with chunk-based processing and a Sparser Branch with CSC inputs and weight forwarding. Empirical results show dramatic speedups over CPU/GPU baselines and state-of-the-art accelerators (e.g., up to 15286× vs CPU, 294× vs GPU, and notable gains on HyGCN/AWB-GCN), along with maintained or improved accuracy and reduced off-chip bandwidth. The approach presents a practical pathway to scalable, low-latency GCN inference on large graphs and enables more ambitious GCN architectures to be explored in real-world settings.

Abstract

Graph Convolutional Networks (GCNs) have emerged as the state-of-the-art graph learning model. However, it can be notoriously challenging to inference GCNs over large graph datasets, limiting their application to large real-world graphs and hindering the exploration of deeper and more sophisticated GCN graphs. This is because real-world graphs can be extremely large and sparse. Furthermore, the node degree of GCNs tends to follow the power-law distribution and therefore have highly irregular adjacency matrices, resulting in prohibitive inefficiencies in both data processing and movement and thus substantially limiting the achievable GCN acceleration efficiency. To this end, this paper proposes a GCN algorithm and accelerator Co-Design framework dubbed GCoD which can largely alleviate the aforementioned GCN irregularity and boost GCNs' inference efficiency. Specifically, on the algorithm level, GCoD integrates a split and conquer GCN training strategy that polarizes the graphs to be either denser or sparser in local neighborhoods without compromising the model accuracy, resulting in graph adjacency matrices that (mostly) have merely two levels of workload and enjoys largely enhanced regularity and thus ease of acceleration. On the hardware level, we further develop a dedicated two-pronged accelerator with a separated engine to process each of the aforementioned denser and sparser workloads, further boosting the overall utilization and acceleration efficiency. Extensive experiments and ablation studies validate that our GCoD consistently reduces the number of off-chip accesses, leading to speedups of 15286x, 294x, 7.8x, and 2.5x as compared to CPUs, GPUs, and prior-art GCN accelerators including HyGCN and AWB-GCN, respectively, while maintaining or even improving the task accuracy. Codes are available at https://github.com/RICE-EIC/GCoD.

Figures (12)

• Figure 1: An overview of the proposed GCoD, an algorithm and accelerator co-design framework dedicated to GCN acceleration.
• Figure 2: Illustrating (a) GCoD's defined group, class, and subgraph within GCNs' graph adjacency matrices, where nodes with similar degrees are categorized into the same class, each class is further divided into subgraphs with a similar number of edges, and all the subgraphs within the same class are evenly distributed into different groups, and (b) each hardware chunk (i.e., sub-accelerator) handles the same kind of classes from all the groups.
• Figure 3: GCoD training flow.
• Figure 4: Visualizing three datasets' graph adjacency matrices before and after applying our GCoD algorithm, where green lines separate subgraph classes while red lines partition the groups. Note that non-zero dots in the matrices are enlarged for better visualization effects.
• Figure 5: Illustrating the gathered and distributed aggregations.