Maze routing is usually the most time-consuming step in global routing or detailed routing. One possible way to accelerate it is to use parallel computing. Net-level parallelism is commonly used but it is affected greatly by the dependency between nets. There are few GPU-friendly parallel maze routers, which can be nontrivial to design. In this paper, we propose a pathfinding-level parallel 3D routing scheme. We implemented it in CUDA and applied it to the coarsened maze routing stage of an open source global router CUGR. Compared with CUGR on the ICCAD 2019 global routing contest benchmark suite, we achieve an average of 16 x speedup in the coarsened maze routing stage without loss of quality.