How RTGI Benchmarks Stress Tensor and RT Cores in GPUs

This article provides a comprehensive overview of how Ray-Traced Global Illumination (RTGI) pushes modern graphics cards to their absolute limits during hardware benchmarks. It explains the fundamentals of RTGI, details how the technology saturates RT cores through complex light-bounce calculations, and examines how Tensor cores are simultaneously maxed out via AI denoising and upscaling. Finally, it covers why the concurrent utilization of these specialized cores makes RTGI the ultimate stress test for evaluating GPU thermal limits, power delivery, and architectural efficiency.

What is Ray-Traced Global Illumination (RTGI)?

Traditional games and benchmarks use rasterization, a technique that fakes lighting using pre-baked data. Ray-Traced Global Illumination (RTGI), however, simulates the actual physical behavior of light in real time. When light hits a surface in the real world, it bounces, illuminating surrounding areas and picking up the color of the surface it hit. RTGI calculates these indirect light bounces dynamically. Because simulating millions of light rays bouncing multiple times per frame requires immense computational power, hardware benchmarks use RTGI to intentionally bottleneck and test the maximum capabilities of a GPU.

How RTGI Stresses RT Cores

RT (Ray Tracing) cores are dedicated hardware units designed to handle the specific math required to trace the path of light. While basic ray tracing might only calculate a single bounce for a reflection or a shadow, RTGI calculates multiple bounces for every light source in a scene. This stresses the RT cores in two primary ways:

• BVH Traversal: To know what a ray of light hits, the GPU breaks the 3D scene down into a complex grid of boxes called a Bounding Volume Hierarchy (BVH). RT cores must continuously search through this hierarchy to trace the path of every single ray. In an RTGI benchmark, objects are often moving, forcing the GPU to rebuild and traverse the BVH simultaneously at lightning speeds.
• Ray-Triangle Intersection Tests: Once a ray reaches the end of the BVH, the RT core must calculate the exact pixel on the 3D model's polygon (triangle) the light hits, and at what angle. Because RTGI requires primary, secondary, and tertiary light bounces to accurately light a room, the number of intersection tests increases exponentially, driving RT core utilization to 100%.

How RTGI Stresses Tensor Cores

Even the most powerful RT cores cannot cast enough rays in real time to create a clear image; doing so would reduce the benchmark to a slideshow. Instead, the GPU casts a limited number of rays, which results in a grainy, pixelated, and incomplete lighting map. This is where Tensor cores are pushed to their limits.

Tensor cores are specialized processors built for high-speed matrix multiplication, which is the foundation of deep learning and artificial intelligence. During an RTGI benchmark, Tensor cores handle two massive workloads:

• AI Denoising: Tensor cores run complex neural networks to "connect the dots" of the grainy ray-traced image. Technologies like Ray Reconstruction use Tensor cores to analyze the sparse ray data alongside previous frames, filling in the missing light information to produce a smooth, photorealistic image in milliseconds.
• AI Upscaling: Because RTGI is so demanding, benchmarks often utilize technologies like DLSS (Deep Learning Super Sampling) or XeSS. The GPU renders the scene at a lower resolution to save performance, and the Tensor cores use AI to upscale the image to the target resolution (like 4K).

The Ultimate Combined Stress Test

In standard gaming workloads, different parts of the GPU often wait on one another, meaning the hardware rarely hits its absolute theoretical limit. RTGI benchmarks are prized by hardware testers because they force the entire GPU architecture to work in tandem at maximum capacity.

While the standard shader cores are rendering the geometry of the scene, the RT cores are maxed out calculating light bounces, and the Tensor cores are fully saturated cleaning and upscaling the final image. Furthermore, storing the massive BVH structures and high-resolution frame data pushes the Video RAM (VRAM) bandwidth to the brink.

Because all three core types (Shader, RT, and Tensor) and the memory subsystem are operating at peak load simultaneously, an RTGI benchmark exposes any weaknesses in a GPU's power delivery, voltage regulation, and thermal cooling solutions faster than almost any other workload.