TL;DR
This paper introduces GenGNN, a novel FPGA acceleration framework designed for a wide range of Graph Neural Networks (GNNs). It achieves significant speedups for real-time inference without graph preprocessing, making it highly adaptable to new and emerging GNN models.
What Problem Does This Paper Address?
Graph Neural Networks (GNNs) are becoming increasingly popular for tasks involving graph-structured data, from drug discovery to high-energy physics. However, deploying these models efficiently, especially for real-time applications, is a major challenge. Developing hardware accelerators for GNNs is complex, and the rapid evolution of GNN architectures means that specialized accelerators quickly become outdated. Existing solutions often focus on specific GNN types (like GCNs) or require extensive preprocessing of the graph data, which is not feasible for real-time scenarios. This creates a gap between the demand for fast GNN inference and the ability to provide generic, adaptable, and efficient hardware solutions.
Methods
GenGNN tackles this problem by proposing a generic framework implemented using High-Level Synthesis (HLS) on FPGAs. The core of GenGNN is an optimized message-passing architecture that is designed to be compatible with a broad spectrum of GNN models. This generic structure is complemented by a rich library of model-specific components, allowing for easy adaptation to new GNN architectures.
Key methodological contributions include:
- Generic Message-Passing Architecture: A unified approach to handle the core computations of GNNs, making it adaptable to various GNN families.
- Model-Specific Components: A modular library that can be extended to incorporate specialized operations for different GNN models (e.g., attention mechanisms, different aggregation functions).
- Real-time Inference without Preprocessing: GenGNN is designed to process raw graph data directly, eliminating the need for time-consuming preprocessing or partitioning steps, which is crucial for real-time applications.
- Streaming-Based Pipelining: An efficient pipelining strategy for node embedding and message passing stages, minimizing idle cycles and maximizing throughput. This is particularly beneficial for models with virtual nodes.
- High-Level Synthesis (HLS): Using HLS allows for faster development and easier adaptation of the framework compared to traditional hardware description languages.
Datasets
The paper evaluates GenGNN using a combination of datasets:
- Molecular Property Prediction: MolHIV and MolPCBA from the Open Graph Benchmark (OGB) were used for evaluating real-time performance on graph-level tasks.
- Node-Level Classification: Cora, CiteSeer, and PubMed were used for evaluating the large graph extension capabilities.
Key Insights and Ideas
- Generality is Key: The paper strongly argues that a generic framework is essential for keeping pace with the rapidly evolving GNN landscape, rather than focusing on specialized accelerators for individual GNN types.
- Real-time without Preprocessing is Achievable: GenGNN demonstrates that it's possible to achieve high-performance GNN inference on FPGAs without the often-required graph preprocessing, opening doors for real-time applications.
- Pipelining for Efficiency: The streaming-based pipelining approach is a critical enabler for efficient execution, especially for complex GNN models and those employing techniques like virtual nodes.
- FPGA's Sweet Spot: FPGAs, with their reconfigurability and parallelism, are well-suited for accelerating GNNs, particularly when customizability and low latency are paramount.
Key Results
GenGNN was implemented on a Xilinx Alveo U50 FPGA and achieved significant speedups:
- Real-time Performance: Up to 25x speedup compared to a CPU baseline (Intel Xeon Gold 6226R) and 13x speedup compared to a GPU baseline (NVIDIA RTX A6000) on molecular property prediction datasets.
- Large Graph Extension: For larger graphs that don't fit on-chip, GenGNN achieved 1.49-1.95x speedup over CPU and 1.32-2.44x speedup over GPU on datasets like Cora, CiteSeer, and PubMed.
- Pipelining Effectiveness: The streaming-based pipelining technique showed speedups of up to 1.92x over non-pipelined approaches and 1.37x over fixed pipelining.
Limitations and Future Work
- Large Graph Optimization: While GenGNN demonstrates scalability for larger graphs, the authors acknowledge that further optimizations are needed to hide memory access latency for graphs that do not fit entirely on-chip.
- Quantization: The current implementation uses 32-bit fixed-point quantization for the Large Graph Extension, and future work could explore 16-bit quantization for further resource savings.
- Design Automation: The authors suggest future work in design automation, design space exploration, and further optimization for large graphs.
- Open-Source Contribution: The HLS code for GenGNN will be open-sourced upon acceptance, fostering community development and further research in this area.
Summary updated