Title: Hardware-software Codesign for accelerating Graph Neural Networks on FPGA
PhD Candidate: Bingyi Zhang
Committee Members: Prof. Paul Bogdan, Prof. Rajgopal Kannan, Prof.
Viktor Prasanna (Chair), Prof. Weihang Wang
Date: Monday, July 1st, 2024
Time: 2 pm
Location:
Zoom Link: https://usc.zoom.us/j/4520579668?pwd=eXNyaThLeFloNzBLVHlZQ0FYdzRGdz09
Abstract: Graph Neural Networks (GNN) have revolutionized many real-world applications, such as recommendation systems, social networks, etc. However, current GNN libraries on general-purpose processors achieve sub-optimal performance due to several challenges: 1. Irregular data structure: the graphs in real-world applications are highly unstructured, with uneven degree distribution. Such irregularity leads to complex data access patterns. 2. Heterogeneous computation kernels: GNNs involve both sparse computation kernels and dense computation kernels. While general-purpose processors are efficient for dense computation, their data path and memory hierarchy are inefficient for sparse computation. 3. Dynamic data sparsity: In many applications, the graph connectivity and the data sparsity of vertex features are unknown before executing the GNN model. Such dynamic data sparsity makes it difficult for the compiler and runtime system to generate an optimal execution scheme for GNN. 4. Mixture of Models: Some GNN-based applications, such as GNN-based computer vision tasks, utilize a mixture of CNN and GNN models. Such a combination leads to complex data flow. In this dissertation, we address the above challenges through novel hardware-software codesign. First, to address the first two challenges, we develop an accelerator-compiler codesign on FPGA for GNN inference, named GraphAGILE, for the end-to-end acceleration of GNNs. Second, we propose Dynasparse, an efficient codesign of runtime system and hardware to exploit the dynamic sparsity in GNN inference. Third, we propose GCV-Turbo, a hardware-software codesign accelerating GNN-based computer vision (CV) tasks, which involves a mixture of GNN layers and CNN layers. Our codesigns achieve superior performance on various GNN-based applications compared with state-of-the-art graph machine learning libraries and hardware accelerators.
Bio: Bingyi Zhang is a fifth-year PhD candidate in Computer Engineering, advised by Professor Viktor K. Prasanna. He received the BS degree in microelectronics from Fudan University in 2017, and the MS degree in Integrated Circuit Engineering from Fudan University in 2019. His research interests include parallel computing, digital signal processing, digital circuit design.
Published on June 10th, 2024Last updated on June 10th, 2024