Using Half-Precision Floating-Point Numbers for Storing Bounding Volume Hierarchies

Bounding volume hierarchies, which are commonly used to speed up ray tracing, are heavily accessed during ray traversal. Reducing the memory footprint for the bounding volume hierarchies leads to a better cache hit ratio and, therefore, faster ray tracing. This paper examines an approach of using 16-bit half-precision floating-point numbers to store bounding volume hierarchy inner nodes. Compared to prior work on BVH node compression, which uses integers, floats are interesting because of their dynamic range. Compared to single-precision, half-precision inner nodes have half
memory footprint and, therefore, they save an average of 1.7% total power. This comes with the small conversion cost overhead, average of 2.5% more ray-box intersection tests and average of 15% more ray-triangle intersection tests. If half floats are used hierarchically, rather than simply storing world coordinates as halfs, roughly the same accuracy is achieved as with single-precision. This eliminates the additional intersection tests, but it requires additional computations to obtain the world coordinates during traversal.