This paper presents a low-power, real-time Simultaneous Localization and Mapping (SLAM) processor which generates fully-dense 3D map through neural-rendering. It is optimized through 3-level hierarchical sparsity exploitation architecture: 1) 3D Sampling Unit (3D-SU) supports sample-level sparsity on 3D space, 2) Sparse Mixture-of-Experts (SMoE) cluster leverages expert-level sparsity of multi-layer perceptron (MLP), and 3) Heterogeneous Coarse-Grained Sparse Core (HCG-SC) handles neuron-level sparsity across three training stages of the MLP. As a result, the processor consumes average 303.5 mW power while achieving over 30 frame-per-second (fps) for end-to-end camera pose estimation and dense RGB-D mapping.