wa-hls4ml: Benchmark and Surrogate Models for hls4ml Resource Estimation
Published in ACM Transactions on Reconfigurable Technology and Systems (TRETS), 2025
We introduce a comprehensive benchmark and dataset containing over 680,000 synthesized neural networks for evaluating surrogate models that predict FPGA resource usage and latency. This work explores graph neural network (GNN) and transformer-based approaches to accurately estimate hardware metrics for hls4ml-synthesized models, enabling faster neural architecture search and hardware-software codesign without requiring full synthesis.