NeuralGI : multi-GPU real-time approximate global illumination using transformer-based AI model

Abstract

We introduce a novel transformer-based global illumination (GI) framework designed to harness underutilized integrated GPUs (iGPUs) alongside discrete GPUs (dGPUs) for real-time lighting in 3D game engines. Our method voxelizes 3D scenes and feeds sequences of per-voxel features into a compact, four-layer transformer model to predict indirect lighting effects such as diffuse light propagation and ambient soft shading. This approach enables visually plausible GI with minimal overhead, operating as a distributed system across Unity and Godot engines. In this architecture, Unity performs transformer-based inference using the iGPU, while Godot handles voxelization and final rendering using the dGPU. Communication between engines is facilitated through memory-mapped files, ensuring low-latency data exchange.

Experimental evaluations on mainstream consumer hardware demonstrate that the system achieves peak performance at a batch size of 104 voxels, delivering over 500 voxel predictions per second with stable GPU memory usage. Qualitative comparisons against ray-traced ground truth and voxel cone tracing (VXGI) show the model's ability to approximate indirect lighting across both familiar and unseen environments. By leveraging transformer architectures for spatial reasoning and effectively utilizing idle compute resources, this work presents a practical step toward integrating machine learning-based GI into real-time game pipelines. It further establishes iGPUs as viable co-processors for lightweight neural rendering tasks in modern heterogeneous computing environments.