New blog post! "Measuring acceleration structures", in which we will compare BVH costs on various GPU architectures and drivers and attempt to understand the details enough on AMD hardware to make sense of the numbers!
Reposts appreciated :)
https://zeux.io/2025/03/31/measuring-acceleration-structures/
Reposts appreciated :)
https://zeux.io/2025/03/31/measuring-acceleration-structures/
Comments
https://projects.blender.org/blender/blender/issues/136174
> the bottleneck is mainly with temporary buffers that are allocated during bvh construction (upward of 10x memory over balanced).
Note that in my post I only discuss final - built and compacted - sizes. Temporary sizes are a separate problem, and are indeed much larger.
On radv, you can use RADV_DEBUG=emulate_rt which disables HW units; this was something like 2-3x slower on 7900.