Measuring Acceleration Structures

Summary

Measuring acceleration structures 31 Mar 2025 Hardware accelerated raytracing, as supported by DirectX 12 and Vulkan, relies on an abstract data structure that stores scene geometry, known as “acceleration structure” and often referred to as “BVH” or “BLAS”. Unlike geometry representation for rasterization, rendering engines can not customize the data layout; unlike texture formats, the layout is not standardized across vendors. It may seem like a trivial matter - surely, by 2025 all implementations are close to each other in memory consumption, and the main competition is over ray traversal performance and new ray tracing features? Let’s find out. Experimental setup It’s going to be difficult to make any generalized claims here; and testing this requires using many different GPUs by many different vendors, which is time consuming. So for the purpose of this post, we will just look at a single scene - Amazon Lumberyard Bistro, or more specifically a somewhat customized variant by Nvidia which uses more instancing than the default FBX download. The results are captured by running niagara renderer; if you’d like to follow along, you will need Vulkan 1.4 SDK and drivers, and something along these lines: git clone https://github.com/zeux/niagara --recursive cd niagara git clone https://github.com/zeux/niagara_bistro bistro cmake . && make ./niagara bistro/bistro.gltf The code will parse the glTF scene, convert the meshes to use fp16 positions, build a BLAS for every mesh, compact it using the relevant parts of VK_KHR_acceleration_structure extension, and print the resulting compacted sizes. While a number of levels of detail are built as the scene is loaded, only the original geometry makes it into acceleration structures, for a total of 1.754M triangles. The builds are using PREFER_FAST_TRACE build mode; on some drivers, using LOW_MEMORY flag allows to reduce the BLAS size further at some cost to traversal performance, which we will ignore for now. Experimental results...