On-board preprocessing is a crucial step in satellite image processing, particularly when implementing edge computing with neural networks. However, this approach presents multiple challenges due to resource limitations at the edge. A key step in overcoming these challenges is selecting an appropriate neural architecture that balances performance with computational constraints. This paper proposes a method to support the decision-making process by defining a Performance Relation Parameter (PRP) for evaluating the performance of a given neural architecture on a specific hardware platform. To demonstrate its effectiveness, we apply this parameter to the task of object detection on a satellite board. We analyze transformer-based architectures in the experiments and compare their backbones using theoretical metrics and actual performance. This comparison provides insights into the trade-off between the models’ and edge hardware’s complexity and effectiveness. The results show that the proposed parameter can be used to identify backbones that achieve high inference efficiency while maintaining accurate predictions. The PRP can also be used as a tool for recommending the most promising models for further optimization and deployment on edge devices.