Work

High-Performance and Flexible Graph Neural Network Inference Acceleration with a FPGA-based Software Programmable Overlay

Year: 2025
Type: dissertation
Abstract: Graph neural networks (GNNs) are a class of deep learning (DL) models for learning latent representations of graph-structured data for a variety of node/graph-level prediction tasks, some of which require real-time low latency inference. Field-Programmable Gate Arrays (FPGAs) are one type of specialized accelerator that can offer significant performance and energy benefits versus traditional central processing units (CPUs). Most existing FPGA-based GNN accelerators prepro cess input graphs, making them unsuitable for low-latency use cases, or are extremely specialized streaming pipelines that require a long FPGA bitstream recompile to target each type of GNN. We enhance the neural processing unit (NPU) overlay architecture, instruction set, and compiler to enable high-performance and flexible acceleration for a variety of GNN models. Compared to prior work, our NPU achieves 7.8× more throughput vs. a model-specific dataflow accelerator and 2.6× lower latency a state-of-the-art Versal processor array accelerator.
Language: English
Open Access status: green
OpenAlex ID: W7139450502
Cites: 92
Cited by:
Related to: 10