Fully Fused Neural Network for Radiance Caching in Real Time Rendering
, NVIDIA
We'll cover a "Megakernel" neural network that has much greater inference and training performance (5-10x) than an equivalent implementation in TensorFlow or PyTorch, enabling its use in real-time rendering. We present a proof-of-concept application of the Megakernel neural network to real-time radiance caching. Thanks to its good performance, this neural radiance cache can be trained online during rendering, adapting quickly to animated content and camera movement while simultaneously offering the high approximation quality that neural networks are known for. The neural radiance cache represents a new point on the spectrum of techniques trading off quality for performance; it can be thought of as a more accurate, albeit more expensive version of RTXGI.