Hi everyone.

Does anyone know, or heard of, exactly what kind of ray tracing hardware acceleraton will GPU vendors provide?

Is it just more shader cores to run compute-like implementation, something like DXR Fallback Layer provides, or things like triangle intersection would be implemented as dedicated fixed-function hardware? Or maybe it is entirely vendor-specific?

NV is the only one to have announced special HW yet (besides PowerVR, who did this first), but the others will surely follow suit.

NV started by adding more flexible thread scheduling capabilities for compute, which is important as shaders can now spawn single rays at any point in time (which can in turn generate single hit-shader invocations, which can spawn more rays, etc) and all these rays/threads will need to be dynamically grouped into batches of at least 32 by the hardware for it to be at all efficient. The details on this are scarce, but some kind of hardware solution to this problem exists in Volta.

Then, NV's Turing includes "RT cores", which are fixed function ray-vs-triangle (and likely ray-vs-box) testing hardware. DXR allows the driver to construct a BVH for your triangles in a way that's hidden from you. This fixed function hardware can apparently perform BVH traversal (lots of recursive ray-vs-box tests) as well as actual ray-vs-triangle tests in the leaves of the BVH. When hit(s) are found, they will queue up hit-shader invocations on the regular compute cores.

1 hour ago, Hodgman said:

which are fixed function ray-vs-triangle

Thanks for the answer. Where did you get that information?
Also, do you mean Turing, not Tesla?