By using this site, you agree to our Privacy Policy and our Terms of Use. Close
JEMC said:
WoodenPints said:

Are we actually expecting much from the new DXR features for RT and the neural shading to lower vram usage or is it just something that sounds great but never really materializes?

I know way too little to know.

Past Direct X APIs were supposed to come with features that would bring great performance gains, but devs didn't use them because they were hard and time consuming to implement.

We can only wait and see what happens, or hope that someone more knowledgeable comments.

There are significant efficiency gains to be had in the upcoming Direct X Ray Tracing, some of it requires developers to support it..


***********************

Opacity micro-maps (OMM) is one such aspect that requires developers to actively seek and use it as it changes the way we traditionally do shadows, fences, foliage and other alpha-tested geometry.

Essentially the calculations are now done before it hits the shaders, which should mean shaders do less work due to more efficient culling.

Basically the RT cores check if a ray intersects the geometry or not.
Without Opacity Micro-Maps, when geometry that uses alpha-maps is hit by a Ray, we don't know if the Ray hit a part that was opaque or transparent, so normally the shaders are leveraged to do those calculations, which eats up a ton of performance.

************************

Shader Execution Reordering (SER) allows you to sort and re-order shaders to calculate together as a group.
A shader unit is a very simple processor that typically performs various degrees of floating point or integer math calculations of various precision's, but GPU's can have thousands of them.

Where things get tricky is that some code paths will have branching results... And those branches need to be calculated before unneeded results are discarded.
So the idea is to try and group shaders with similar/same results for more efficient execution so that multiple sets of branch calculations don't need to be calculated.


************************

Cooperative Vectors (CV) is just essentially hardware support for new Vector/Matrix math.
This will enable things like Neural Block Texture Compression, enhanced Neural Supersampling and Denoising.

Even nVidia's new Neural Shading will support this.

***********************

Keep in mind that we are still in the infancy of hardware accelerated Ray Tracing and A.I... It's been a path that industry has been working towards for decades, there are significant gains to be had going forwards.

Hopefully this answers some of your questions/unknowns.




www.youtube.com/@Pemalite