NeRPIM: A 4.2 mJ/frame Neural Rendering Processing-in-memory Processor with Space Encoding Block-wise Mapping for Mobile Devices

Abstract

NeRPIM, a world-first energy-efficient 3D neural rendering processing-in-memory (PIM) processor, is proposed for mobile devices. A PIM-based renderer reduces the per-sample rendering energy with 8T SRAM PIM. Space encoding blockwise mapping reduces the global memory access and reduces overall rendering energy. A ray-wise dynamic block reuse maximizes the reusability of PIM with enhanced throughput. NeRPIM is fabricated in 28nm CMOS technology and occupies 5.1 mm <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of die area. As a result, the processor only consumes 129.8 mW of power at 30.7 FPS (NeRF synthetic dataset), achieving a state-of-the-art energy efficiency of 4.2 mJ/frame.