NRX: An End-to-End Real-Time Neural Rendering Pipeline Optimized for Modern AI GPUs

Real-time neural global illumination (GI) on modern AI GPUs remains challenging: existing systems optimize single components (e.g., 3D Gaussian splatting or a GI network) and leave pipeline-level opportunities in precision, tiling, and scheduling unexploited. We present NRX, an end-to-end real-time neural rendering pipeline for high-quality GI. NRX combines a hybrid mesh--Gaussian representation with a lightweight neural network stage that predicts multi-bounce transport from fused G-buffers and Gaussian features. The pipeline runs entirely on the GPU, starting from mesh-based soft rasterization. Our core contribution is pipeline-level orchestration---coordinating precision-aware execution, tile-aware scheduling, and architecture-conscious optimizations across heterogeneous stages. NRX applies FP16/BF16 (and optional FP8) driven by stage-wise numerical sensitivity, jointly partitions mesh and Gaussian rendering into spatial tiles to alleviate workload skew, and employs kernel fusion, data-layout tuning, and stream-level overlap. Experiments on H100 and RTX~5090 across three representative scenes show 18--23% (H100) and 20--30% (RTX~5090) end-to-end speedup over FP32 baseline with negligible quality degradation (△SSIM≤0.0081, △PSNR≤0.25 ,△LPIPS≤0.0056).