上海科技大学知识管理系统

Photorealistic rendering of virtual scenes is a crucial task in the applications of augmented reality (AR), virtual reality (VR) and metaverse. Traditional rasterization or raytracing-based approaches usually involve a lot of manual efforts, making them unsuitable for mass production of virtual contents. The recently proposed Neural Radiance Fields (NeRF)-based technique demonstrated state-of-the-art results, but it demands a huge amount of computation. In this paper, we propose a resistive random access memory (RRAM)-based NeRF rendering accelerator. In the proposed accelerator, dataflow is specifically designed according to the parallel nature of NeRF-based rendering to increase the utilization of the RRAM array. Moreover, positional encoding engines are included in the proposed accelerator to reduce the off-chip memory bandwidth. Simulation results show that, for the execution of NeRF-based rendering, the overall performance and energy efficiency of the proposed accelerator is much higher than the existing GPU and TPU implementations. © 2022 IEEE.