A Low-Power Neural Graphics System for Instant 3D Modeling and Real-Time Rendering on Mobile AR/VR Devices

This work presents a low-power neural graphics system enabling both instant 3D modeling and real-time rendering on mobile devices with 3 key features: 1) Segmented Hashing with Spatial Pruning (SHSP), 2) Hybrid Interpolation Unit (HIU), and 3) Similarity-Sparsity Skipping (S
³
) Core. Resolving irregular memory access by hash function and high computation cost of MLP, proposed system achieves 8.7× faster 3D modeling and 231.4× smaller energy per iteration compared to edge GPU although it has 31× smaller off-chip memory bandwidth.