Abstract
As the size of real-world graphs increases, training Graph Neural Networks (GNNs) has become time-consuming and requires acceleration. While previous works have demonstrated the potential of utilizing FPGA for accelerating GNN training, few works have been carried out to accelerate GNN training with multiple FPGAs due to the necessity of hardware expertise and substantial development effort. To this end, we propose HitGNN, a framework that enables users to effortlessly map GNN training workloads onto a CPU-Multi-FPGA platform for acceleration. HitGNN consists of high-level APIs, a software generator, and an accelerator generator with optimized datapath and memory organization. Compared with state-of-the-art multi-GPU platforms, HitGNN achieves up to 27.21x bandwidth efficiency and up to 4.26x speedup while demonstrating good scalability to 16 FPGAs.
Introduction
Training Graph Neural Networks (GNNs) on massive real-world graphs presents a significant computational challenge. While GPUs are the standard workhorse, they often struggle with the irregular, sparse memory access patterns inherent in graph data, leading to poor memory bandwidth utilization. FPGAs (Field-Programmable Gate Arrays) offer a promising alternative because their hardware can be customized to match specific data access patterns. However, deploying GNN training on multi-FPGA systems is notoriously difficult due to the required hardware expertise, the complexity of workload balancing, and the development effort needed to map complex algorithms like DistDGL or PaGraph to hardware.
HitGNN bridges this gap by automating the deployment of GNN training onto heterogeneous CPU+Multi-FPGA platforms. By combining high-level APIs with automated hardware generation and smart workload balancing, it achieves significant performance gains over state-of-the-art multi-GPU systems.
Methods
The HitGNN framework is built upon three main pillars designed to abstract hardware complexity while maximizing performance:
- High-Level APIs: Users define GNN models (e.g., GCN, GraphSAGE) and training algorithms using Python-like APIs, requiring only a few lines of code.
- Automated Generators:
- Software Generator: Creates the host program to manage mini-batch sampling, CPU-FPGA communication, and workload balancing.
- Accelerator Generator: Utilizes a Design Space Exploration (DSE) engine to automatically determine optimal hardware parameters (e.g., processing elements) based on available FPGA resources.
- Multi-FPGA Optimizations:
- Two-Stage Task Scheduling: Implements a dynamic scheduler to redistribute mini-batches from busy to idle FPGAs, preventing workload imbalance.
- Direct CPU Data Fetching: Leverages the host CPU's large memory as a central hub. FPGAs pull missing vertex features directly via PCIe, reducing the overhead associated with complex chip-to-chip communication.
Key Insights and Results
The evaluation of HitGNN on large-scale datasets like Reddit, Yelp, Amazon, and ogbn-products revealed several critical insights. First, customization often outweighs raw peak power; despite having lower peak TFLOPS than high-end GPUs, HitGNN's hardware-specific optimization for GNN memory access patterns leads to higher effective performance. Second, the framework's DSE engine identifies and mitigates shifting bottlenecks between the "Aggregation" and "Update" steps of GNN layers.
Empirical results show that HitGNN achieves a geometric mean speedup of 2.11× to 2.34× over a 4-GPU baseline, with peak speedups reaching 4.26×. Notably, it demonstrates up to 27.21× higher bandwidth efficiency, performing more work with significantly less memory bandwidth. The system also exhibits near-linear scalability up to 16 FPGAs, proving its capability for industrial-scale graph processing.
Conclusion
HitGNN successfully automates the deployment of GNN training on CPU+Multi-FPGA platforms, overcoming traditional barriers of hardware expertise and development effort. While the current implementation identifies CPU memory bandwidth as a potential future bottleneck (around 12-16 FPGAs), future work in data prefetching and high-speed interconnects (like SmartNICs) could further extend these performance gains.
Summary updated