Many image super-resolution networks with arbitrary upscale factors has been proposed and shown remarkable performance. In this work, we employ SR neural network to reconstruct power distribution of an assembly in reactor from a coarser mesh to arbitrarily finer mesh. Based on SRNO, we rebuild it to a 3D version SRNO-3D, and add output normalization at the end of the model. This optimization has been proved as a cheap but powerful method to enhance reconstruction accuracy, which is inspired by the prior knowledge that power tally data is indicative of relative distribution. In addition, we test the model with continuous in-distribution SR and out-of-distribution upscaling, the quantitative results reveal unstable performance, especially for radial SR, which demonstrates that 3D power reconstruction is obviously affected by physical constrains such as the lattice structure and zero power of non-fuel areas. This work validates the feasibility of 3D arbitrary upscale SR of power distribution for the first time, and points out the challenge for future study.