A recent development in mapping machine learning models onto reconfigurable dataflow hardware introduces a learned cost model designed to address the challenges of throughput evaluation. Traditional methods often depend on intuition, which can lead to inaccuracies in predicting throughput performance. The learned model improves upon these approaches by providing more accurate throughput predictions, as supported by current evidence. This enhanced prediction capability has the potential to increase the efficiency of dataflow-graph mappings on such hardware. By refining throughput evaluation, the model could optimize resource allocation and execution strategies in reconfigurable systems. The innovation represents a promising step toward more effective deployment of ML models in hardware environments that require dynamic reconfiguration. Overall, this approach may contribute to advancing the performance and adaptability of dataflow-based machine learning accelerators.