Hyperspectral remote sensing applications have been greatly increased in the last decades and are expected to continue growing in number in the near future. In timecritical scenarios, such as target detection for military purposes, monitoring of chemical contamination, or wildfire tracking, it is important to accelerate hyperspectral data analysis techniques. In order to avoid high data volume transmission bottlenecks and the associated delays, onboard processing is highly desirable. The recent deep learning models have shown very good results on image classification and therefore have been applied also in hyperspectral image classification. However, running these algorithms in low-power on-board devices is a challenging task since they require high computing power and memory. Architectural optimizations and performance accuracy tradeoffs must be established to find a computing solution that guarantees the real-time analysis of hyperspectral images. This paper analyzes the performance accuracy tradeoff of a deep learning model for hyperspetral image analysis and classification. Changing the characteristics of the model and quantizing its parameters leads to different ratios between performance and accuracy when implemented in FPGA. The results show that the computational solution can be considerably improved with negligible accuracy reduction.