In an unexpected challenge for today’s top-tier GPUs, a new DirectX 12 benchmark has emerged that forces even the GeForce RTX 5090 to its computational knees—even when rendering at just 480p. The Radiance benchmark, developed by former Toms Hardware and Thresh Alan Dong, doesn’t rely on traditional ray tracing or texture mapping but instead uses raymarching, a method that calculates light paths purely through mathematical functions. This approach, while visually impressive, demands an unprecedented level of FP32 compute throughput, leaving current hardware struggling to maintain playable frame rates without significant compromises.
The benchmark’s design is intentionally minimalistic, with a compressed size of just 80KB and two presets: one set at 720p resolution with 80 debris elements, and another extreme preset pushing 1080p with 640 debris elements. Despite its small footprint, the test places extreme demands on GPU execution efficiency, bypassing traditional rasterization in favor of a compute shader-driven approach that marches rays through scenes using signed distance functions (SDFs). This method eliminates the need for pre-baked lighting or texture maps, making it a pure test of raw computational power rather than memory bandwidth or specialized acceleration units.
When tested on an RTX 5090, the benchmark delivered 2085 points in the default preset at 720p, with an average frame rate of 76.2 FPS and debris rendering at 58.3 elements per frame. However, when switching to the extreme preset—where resolution jumps to 1080p and debris count increases to 640—the GPU barely managed 2-3 FPS, with an average debris count of just 123.1. These results underscore the benchmark’s ability to stress-test current hardware while also serving as a glimpse into the kind of computational demands future rendering techniques may impose.
The implications for the industry are significant. Raymarching, though not yet widespread in games, represents an alternative to traditional rasterization and ray tracing that could become more relevant as developers seek ways to achieve photorealistic visuals without relying solely on hardware-accelerated ray tracing. The Radiance benchmark effectively isolates FP32 compute performance, making it a valuable tool for comparing GPUs based on their ability to handle complex mathematical operations—a capability that will likely grow in importance as rendering techniques evolve.
For now, the benchmark serves as both a challenge and a preview. While today’s top-end GPUs like the RTX 5090 can barely eke out playable performance at extreme settings, the test also hints at the kind of computational power that may be required for next-generation rendering techniques. As GPU architectures advance, this benchmark could become a critical reference point for measuring progress in execution efficiency and FP32 throughput.
- Display: 720p (default preset), 1080p (extreme preset)
- Resolution Tested: 720p, 1080p
- Debris Elements: 80 (default), 640 (extreme)
- Ray Steps: 72 (extreme preset)
- GPU Tested: NVIDIA GeForce RTX 5090
The Radiance benchmark is available for download, allowing users to test their own GPUs and share results. While it may not yet be practical for real-world gaming, its ability to push hardware to its limits makes it a compelling tool for understanding the boundaries of current GPU architectures.
