[SIBGRAPI 2026]
A Multi-Engine Benchmark for
Relativistic Raytracing
1 Universidade Federal Rural de Pernambuco (UFRPE), Recife, Brazil
Abstract
Interactive relativistic ray tracing has shifted from offline pipelines to real-time GPU workloads, allowing the visualization of gravitational light deflection via per-pixel geodesic integration. However, developers face critical engineering tradeoffs between high-level engine abstractions (e.g., Unity) and low-level APIs (e.g., Vulkan), as well as the choice of numerical integrators across different spacetime geometries.
This paper introduces A Multi-Engine Benchmark for Relativistic Raytracing, a comprehensive performance-accuracy dataset spanning 16,656 timed renders across three spacetime metrics (Newton, Schwarzschild, Kerr), four integrators, and six scenes of varying multi-body complexity. Our findings reveal a resolution-dependent platform crossover: Unity outperforms Vulkan at ultra-low scales (144p) due to dispatch-overhead dominance, whereas Vulkan achieves up to a 2x throughput advantage at interactive resolutions. We identify the first-order Euler method with a medium step size (h=52 km) as a highly practical sweet spot (>1000fps) for interactive rendering, while demonstrating that extreme multi-body configurations mandate higher-order schemes like RK4.
Interactive WebGL Viewer (144p)
Experience real-time geodesic integration directly in your browser. This custom WebGL implementation runs at a strict 144p resolution (256x144) to match our ultra-low scale benchmarks, utilizing the Euler integrator to visualize Scene 1 (Single Black Hole) and Scene 3 (Accretion Disk) parameters.
Rendering: WebGL 2.0 Compute Shader (Fragment-based)
FPS: 0
Simulation Parameters
Performance & Scaling Results
Platform Crossover
At sub-millisecond GPU kernel durations (144p), Unity's highly optimized engine framework manages command synchronization with tight integration intervals, minimizing driver idling and outperforming Vulkan by up to 1.6x. However, at interactive resolutions (480p+), Vulkan's thinner driver abstraction yields a persistent advantage, especially in computationally dense multi-body scenes.
| Resolution | Newton (Euler) | Schw. (RK4) | Kerr (RK4) |
|---|---|---|---|
| 144p | ~1200 FPS | ~400 FPS | ~300 FPS |
| 480p | ~400 FPS | ~100 FPS | ~75 FPS |
| 1080p | ~120 FPS | ~30 FPS | ~20 FPS |
Integrator Accuracy Tradeoff
The Euler-L to RK4 throughput gap spans 95x (Newton) to 141x (Kerr). The gap widens with metric complexity because RK4's fourfold Christoffel evaluation amplifies Kerr's intensive arithmetic overhead. Below, we present the structural similarity (SSIM) and processing speedup of Euler step configurations relative to the RK4 baseline at 480p (Vulkan).
| Integrator | PSNR (dB) | SSIM (μ±σ) | Speed-up |
|---|---|---|---|
| Euler-L (260km) | 12.4 | 0.47 ± 0.25 | 141x |
| Euler-M (52km) | 15.1 | 0.57 ± 0.26 | 66x |
| Euler-S (1km) | 25.8 | 0.78 ± 0.29 | 2.5x |
Citation
@inproceedings{cavalcanti2026relativistic,
title={Relativistic Raytracer: A GPU Benchmark Across Metrics, Integrators, and Platforms},
author={Cavalcanti, Artur Raffael and Silva Figueiredo, Lucas},
booktitle={39th Conference on Graphics, Patterns and Images (SIBGRAPI)},
year={2026},
organization={IEEE}
}
This project relies on theoretical foundations by Schwarzschild (1916), Kerr (1963), and geometric numerical integration methods.