Constraints of Upper Mantle Thermodynamics with the Very Broadband Rheology Calculator and Magnetotelluric Observations

Abstract

This thesis explores the design of a flexible dataflow architecture to accelerate dynamic Graph Neural Networks (GNNs) for real-time applications requiring strict low-latency processing. Focusing on GNN models based on Edge Convolution (EdgeConv), architectural enhancements are proposed to enable runtime graph construction, dynamic edge embedding inference, and a graph-level pipelined workflow on Field-Programmable Gate Arrays (FPGAs). The design is implemented using Vitis HLS 2024.2 and targets a Xilinx Alveo U50 FPGA platform. Experimental results demonstrate that the proposed architecture achieves an end-to-end inference latency of 62.68 µs, substantially improving over CPU, GPU, and baseline FPGA implementations, and supporting the demands of real-time data processing systems.