Abstract
<title>Abstract</title> <p>Real-time fluid rendering requires balancing visual clarity with computational efficiency, traditionally achieved through manual mathematical shaders. Thiswork investigates whether NVIDIA’s RTX Neural Shaders (RTXNS) framework canreplace conventional vertex displacement and pixel shading logic in real-time fluidrendering pipelines using GPU Tensor Core accelerated neural inferencing. A dualMLP training pipeline was implemented: a vertex shader MLP for mesh-vertex displacement with wave parameter based deformation, and a pixel shader MLP forphysically-based BSDF evaluation. The pixel pipeline achieved successful convergence within seconds, producing visually accurate fluid appearance across diversematerials (water, oils, viscous fluids) at 100 FPS with 60–70 µs per-pixel inferenceon RTX 4060 hardware. However, the vertex pipeline encountered significant framework limitations. Networks replicating the vertex shader produced extremely skewedmeshes at best and blank output at worst. Increasing neurons beyond 32 per hiddenlayer caused drastic slowdowns and application hanging, while importing externallytrained PyTorch weights failed due to proprietary constraints of the .ns.bin formatand absent validation tooling. These findings establish neural BSDF approximationas viable for real-time applications while identifying architectural barriers restricting neural vertex shader deployment. This work assesses RTXNS applicability toreal-time fluid surface rendering and provides concrete directions for framework refinement.</p>