A 0.59μJ/pixel High-throughput Energy-efficient Neural Volume Rendering Accelerator on FPGA

The photorealistic rendering of virtual world has always been a significant challenge in computer graphics. With the advent of Neural Radiance Fields Rendering (NeRF) [1], the past three years has witnessed the explosive development of Neural Volume Rendering (NVR), a data-driven solution to this long-stand problem. However, the scene-representation characteristics of NVR models, coupled with the utilization of Multi-Layer Perceptron (MLP), pose a challenge for the efficient deployment of NVR to existing graphics hardware and dedicated convolutional neural network (CNN) accelerators.