RT-suitedness: predicting hardware ray tracing energy gains from static scene geometry

Official URL: https://dx.doi.org/10.1007/s00371-026-04576-x

Ray tracing has become the standard for photorealistic offline rendering and is increasingly prevalent in real-time applications, yet its energy consumption remains a critical concern, particularly for industrial workloads. This paper presents a comparative study of the energy efficiency of various Bounding Volume Hierarchy (BVH) construction heuristics and traversal methods. We evaluate hardware-accelerated ray tracing against software-based kernels with a specific focus on CAD models. We analyze the trade-offs between traversal latency and total energy cost using an NVIDIA RTX 5090 GPU (Blackwell architecture), supplemented by cross-architecture validation on an RTX 4090 (Ada Lovelace). We demonstrate that optimizing BVH topology for minimal render time directly minimizes dynamic power consumption; consequently, the Surface Area Heuristic (SAH) or its Full-Sweep Variant (FSSAH) combined with tree rotation optimizations [9] yields the highest overall software energy efficiency. Furthermore, we reveal that the energy consumption gains achieved by fixed-function RT cores are strongly modulated by scene topology in ways that can be characterized by measurable geometric properties. To quantify this, we introduce the RT-Suitedness Score (SRT), a novel predictive heuristic that forecasts hardware energy efficiency tiers based on static, measurable geometric properties such as triangle count and variance in primitive aspect ratios (the ratio between triangle edges), without requiring physical render execution.