Atrial Fibrillation (AF) is considered to strongly correlate with stroke. Deep Neural Networks (DNNs) improve the accuracy in real-time atrial fibrillation detection. However, the deployment of DNNs on embedded systems is challenging due to hardware resources. To reduce computation loads, we study the feasibility to eliminate the redundant information in the AF detection task by downsampling. It is compatible with kernel-level optimization, quantization optimization, and model compression methods. A state-of-the-art deep learning model is used to estimate the amount of AF detection information among different sampling rates. This work considers both fixed-length and variable-length time intervals of an Electrocardiograph (ECG) segment. Experiment results demonstrate that model performance can be retained perfectly in AF detection. Ablation study experiments demonstrate the robustness with downsampled signals. Using a large time interval, the AF detection accuracy with 60 Hz signals can be compared to that with 300 Hz signals. Our on-going work includes designing lightweight DNN models with downsampled signals, further exploring the robustness of downsampled signals and model compression.