←back to thread

CUDA Ray Tracing 2x Faster Than RTX: My CUDA Ray Tracing Journey

(karimsayedre.github.io)
59 points ibobev | 2 comments | 25 Jun 25 21:42 UTC | HN request time: 0.451s | source
Show context
sheepscreek ◴[25 Jun 25 22:47 UTC] No.44382488[source]▶
I’m guessing it’s because they’re using all the computing power the GPU has to offer in CUDA mode, as opposed to sharing the GPU with other functions (when in RTX).
replies(2): >>44384359 #>>44384406 #
colechristensen ◴[26 Jun 25 05:13 UTC] No.44384359[source]▶
Yup this is an "assume spherical cow" situation where it's not dishonest, but you can't draw any real world conclusions from the experiment unless you happen to be working in a very restricted space.
replies(1): >>44384383 #
1. ChocolateGod ◴[26 Jun 25 05:18 UTC] No.44384383[source]▶
Wouldn't you need to in a real world scenario make the CUDA cores aware of the game geometry adding more work on the CPU?
replies(1): >>44388850 #
2. touisteur ◴[26 Jun 25 16:20 UTC] No.44388850[source]▶
Ideally you don't make the cuda cores aware but rather the ray-tracing circuitry. RT cores are designed to perform ray-triangle intersections in a BVH. You get the teraflops and memory bandwidth (or more of it) if you fit the RT-core computing model.

And in most cases it's ok to spend time on one CPU function (creating and loading the BVH) against the hundred thousands of frames you'll be drawing on GPU.