In autonomous driving, lidar is inherent for the understanding of the 3D environment. Lidar sensors vary in vertical resolutions, where a denser pointcloud depicts a more detailed environment, albeit at a significantly higher cost. Pointcloud upsampling predicts high-resolution pointclouds from sparser ones to bridge this performance gap at a lower cost. Although many upsampling frameworks have achieved a robust performance, a fair comparison is difficult as they were tested on different datasets and metrics. In this work, we first conduct a consistent comparative study to benchmark the existing algorithms on the KITTI dataset. Then, we observe that there are three common factors that hinder the performance: an inefficient data representation, a small receptive field, and low-frequency losses. By leveraging the scene geometry, a new self-supervised geometric lidar pointcloud upsampling (GLPU) framework is proposed to address the aforementioned limitations. Our experiments demonstrate the effectiveness and superior performance of GLPU compared to other techniques on the KITTI benchmark.