AMD patent sheds light on Radeon's Raytracing plans

The key part here that will reduce the die size/transistor count cost against NVidia's approach is that they can reuse more of the existing registers & caches, these make up a significant portion of modern processors by area and recent node shrinks have had far less impact here than on the logic side of things. They'll almost certainly still have to increase the size of the existing texture caches to be optimal for raytracing but that would benefit traditional shading techniques to a degree too and it ensures minimal dark silicon during non-hybrid rendering(You're usually looking at a 2:1 ratio at least for cache:logic with raytracing hardware).

Also for anyone who's read the patent, when they refer to "fixed hardware raytracing solutions" they mean units like Imagination Technologies RTU's, you could also say NVidia's approach is hybrid in comparison to past attempts at hardware raytracing units like that, though not quite to this degree.
 
Last edited:
The key part here that will reduce the die size/transistor count cost against NVidia's approach is that they can reuse more of the existing registers & caches, these make up a significant portion of modern processors by area and recent node shrinks have had far less impact here than on the logic side of things. They'll almost certainly still have to increase the size of the existing texture caches to be optimal for raytracing but that would benefit traditional shading techniques to a degree too and it ensures minimal dark silicon during non-hybrid rendering(You're usually looking at a 2:1 ratio at least for cache:logic with raytracing hardware).

Also for anyone who's read the patent, when they refer to "fixed hardware raytracing solutions" they mean units like Imagination Technologies RTU's, you could also say NVidia's approach is hybrid in comparison to past attempts at hardware raytracing units like that, though not quite to this degree.

Will be interesting to see what happens under loads that utilize the texture pipeline heavily already that might also have ray tracing now. As you mentioned they will need to increase the size of the traditional texture cache but by how much will be interesting.

there might be certain scenarios that make it fall in a heap

I can picture people cooking up 'special' benchmarks to exploit this potential weakness to highlight perceived shortcomings of the architecture.
 
If you're using raytracing your world's will be stored as voxels in any practical efficient implementation so you shouldn't really be using textures for the main elements being raytraced, so by using RT you naturally decrease texture use essentially. You could concoct some really weird impractical benchmarks to trip it up but I'm not sure anyone would bother.

Using octree-stored voxels and then raycasting basically allows us to treat static 3D geometry as massive texture maps.
 
Back
Top