Cross-Platform Performance Analysis of GNN-Based SAR-ATR on FPGAs

• Frederico Meletti Rappa USP
• Alfredo Goldman USP

DOI: https://doi.org/10.5753/eradsp.2026.30921

Resumo

Graph Neural Networks have been proposed as a lightweight alternative to CNNs for SAR Automatic Target Recognition (SAR-ATR), representing images as sparse graphs to reduce computational cost. This work extends an existing GNN-based SAR-ATR FPGA accelerator with feature aggregation parameterization and cross-platform evaluation on Xilinx Alveo U280 and U55C boards, exploring the impact of SLR mapping, memory configuration, and pipeline parallelism. Results show that compared to the original design single-SLR placement reduces latency and that fewer scatter-gather pipelines result in better performance.

Referências

Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and Dahl, G. E. (2017). Neural message passing for quantum chemistry. In Int. Conf. on Machine Learning, pages 1263–1272. Pmlr.

Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M., and Monfardini, G. (2008a). Computational capabilities of graph neural networks. IEEE Transactions on Neural Networks, 20(1):81–102.

Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M., and Monfardini, G. (2008b). The graph neural network model. IEEE transactions on neural networks, 20(1):61–80.

Wickramasinghe, S., Lin, Y.-C., Raghavendra, C., and Prasanna, V. (2025). Smart: High-performance sar atr through model-architecture co-design on fpga. In 2025 IEEE 33rd Int. Symp. on Field-Programmable Custom Computing Machines (FCCM), pages 75–84.

Zhang, B., Kannan, R., Prasanna, V., and Busart, C. (2022). Accurate, Low-latency, Efficient SAR Automatic Target Recognition on FPGA. In 2022 32nd Int. Conf. on Field-Programmable Logic and Applications (FPL), pages 1–8.

Zhang, B., Kannan, R., Prasanna, V., and Busart, C. (2023). Accelerating GNN-Based SAR Automatic Target Recognition on HBM-enabled FPGA. In 2023 IEEE High Performance Extreme Computing Conf. (HPEC), pages 1–7.