This Fully Raymarched DX12 GPU Benchmark Will Destroy Your RTX 5090 Even at 480p: Based Purely on FP32 Compute & You Can Try It Out Too

A new DX12 GPU benchmark has been launched, which brings even the fastest graphics card, such as the RTX 5090, to its knees.

The Radiance DX12 GPU Benchmark Utilizes Raymarching, Relies Purely on FP32 Compute To Showcase The Potential of Current & Future Graphics Cards

The Radiance benchmark is written by Former Toms Hardware & Thresh's FiringSquad writer, Alan Dang. The benchmark leverages the DX12 API and analyzes the FP32 compute performance of modern GPUs by running a raymarched version of breakout.

Radiance

A Raymarching Benchmark

The same Breakout geometry — bricks, paddle, ball — rendered through pure mathematics. Raymarching calculates the path of every photon, simulating light as it truly behaves. Signed distance functions define surfaces. Global illumination fills the space. Shadows fall where physics dictates they must.

This is our creative vision for what these games could become: the timeless mechanics we've refined, expressed through computational light. The same heritage, the same precision, but rendered with techniques that push current hardware to its absolute limits.

The vision currently exceeds the technology. Radiance requires an RTX 5090 to achieve playable framerates with advanced lighting — and that's with the debris system disabled. Full visual fidelity awaits hardware perhaps a generation away. We release it now as a benchmark: a way to stress-test today's GPUs, and a glimpse of where we're headed.

The fun part is that Alan started Radiance as a weekend project before turning it into a full-on GPU benchmark. The benchmark relies solely on raymarched geometry, no texture maps, no shortcuts, no pre-backed illumination. It's all mathematics!

The benchmark is very compact, with an 80KB compressed size, and has two presets. The first preset called "RTX 5090" has a 720p resolution and features 80 debris, while the Extreme preset runs at a 1080p res & features 640 debris elements. The total size of the default preset is around 5KB, while the Extreme preset is approximately 32KB, which means that even at full capacity, it can fit comfortably within the L1 cache on modern GPUs.

Radiance's best part is that it tests the computational throughput (FP32) and execution efficiency of GPUs, not VRAM bandwidth or other fancy tricks such as RT or AI-accelerated hardware. But the most important aspect about Radiance is its reliance on a rendering method called raymarching. Ray Marching is mainly used in physics simulations as an alternative to ray tracing. The following is the difference between traditional rasterization and raymarching:

Traditional Rasterization

• Vertex Shader transforms coordinates
• Rasterizer determines pixel coverage
• Pixel Shader applies textures/lighting
• Output Merger composites to framebuffer

Raymarching (This Benchmark)

• Compute Shader dispatches 1 thread/pixel
• Each thread marches ray through scene SDF
• On hit: trace shadow rays, calculate lighting
• Write directly to the UAV texture

Alan has a more comprehensive and in-depth guide to Radiance, what it is used for, how it benchmarks the compute resources of GPUs, and what the expectations should be. You can read all here.

So with that said, let's start with benchmarking the NVIDIA GeForce RTX 5090. In the default preset at 720p, we were able to get a score of 2085 points with an average FPS of 76.2 and an average Debris of 58.3.

At the Extreme preset, which is only 1080p with a Debris Cap of 640 and 72 Ray Steps, our RTX 5090 test setup managed just 2-3 FPS with an average Debris size of 123.1. The average FPS was 41.8, but that was before the debris kicked in.

The benchmark is so compute-intensive on the GPU that the developer has added a cautionary note stating that those who run the benchmark ensure good cooling for the graphics card, and also ensure proper cooling and seating of power cables such as 12VHPWR.

Overall, the Radiance benchmark is a nice new test for modern-day GPUs, and we will be keepinga close look at how future GPUs evolve in this test. You can download the benchmark here and share your scores in the comments below.

Follow Wccftech on Google to get more of our news coverage in your feeds.