The Bounding Volume Hierarchy (BVH): How Ray Tracing Calculates Light Paths Without Rasterizing a Scene

The binding constraint of ray tracing is that simulating light requires checking the path of millions of individual rays against the geometry of a 3D scene. If a modern game environment contains ten million polygons, and the camera fires one ray for every pixel on a 4K display, checking every ray against every polygon would require over 80 trillion intersection tests per frame. At 60 frames per second, that brute-force approach is a mathematical impossibility for any silicon that fits inside a console or a desktop PC.[1][3]

To solve this, graphics engineers do not attempt to calculate every intersection. Instead, they rely on an algorithmic shortcut called a Bounding Volume Hierarchy (BVH). A BVH is a tree-like data structure that groups the geometric objects of a scene into nested, conservative bounding boxes. The root node of the tree is a single massive box that encloses the entire rendered world. That box is subdivided into smaller boxes, which are subdivided again, continuing recursively until the smallest boxes—the leaf nodes—contain just a few individual triangles.[1][2]

When a GPU traces a ray of light, it does not test the ray against the scene's geometry directly. It tests the ray against the largest bounding boxes first. If the ray misses a bounding box, the hardware instantly knows it has also missed every object contained within that box. By skipping entire branches of the tree, the BVH reduces the time complexity of finding a collision from linear to logarithmic. Millions of potential polygon checks are discarded in a single mathematical step.[1][2]

This traversal process is so critical to modern rendering that chipmakers have built dedicated silicon specifically to execute it. Nvidia's RT Cores and AMD's Ray Accelerators are specialized hardware units designed to do exactly two things: rapidly navigate the BVH tree and calculate the final ray-triangle intersections at the leaf nodes. By offloading this specific, computationally heavy workload, these units free up the standard shader cores to handle texture mapping and other traditional rasterization tasks.[2]

The architecture of the BVH tree itself is now evolving to match the capabilities of newer hardware. Standard implementations typically use a branching factor of four, known as BVH4, meaning each parent box contains up to four child boxes. However, the custom AMD RDNA architecture powering Sony's PlayStation 5 Pro introduces hardware support for an 8-way Bounding Volume Hierarchy, or BVH8. By doubling the number of child nodes per box, the BVH8 structure creates a shallower, wider tree.[3]

A shallower tree requires fewer steps to traverse from the root to the leaf nodes. For the GPU, this translates directly into fewer memory fetches and faster intersection queries. When CD Projekt Red deployed its PS5 Pro update for Cyberpunk 2077, the studio explicitly cited the new BVH8 implementation as the mechanism that allowed them to enable ray-traced reflections, shadows, and emissive lighting while maintaining playable frame rates. The hardware is simply spending less time searching the data structure.

The practical result of this algorithmic efficiency is visible in the performance metrics of the latest console titles. In recent testing of Marvel's Wolverine running on the PS5 Pro, the advanced ray tracing mode incurred a performance penalty of just 3 to 4 frames per second compared to the standard rendering mode. The console is able to handle complex global illumination and reflection calculations with a negligible hit to overall fluidity.

The push for photorealism in interactive media is not just a matter of adding more raw compute power. It is a continuous refinement of how data is organized and queried. As developers continue to optimize their BVH structures and hardware architects build wider traversal pipelines, the computational cost of simulating physical light will continue to drop, moving the industry further away from the optical illusions of rasterization.[3]