Intel ARC GPUs, Xe Architecture for dGPUs

Lurkmass

Regular
RT shaders still have to be compiled be they compute or RT, so both would cause more PSOs to compile. Compiling "uber-materials" doesn't sound great either.
Separate "RT shaders" (ray generation/intersection/miss/callable or closest/any-hit shaders) don't exist at all with inline RT. We add "RayQuery objects" to either existing compute or pixel shaders that already handle our material and lighting in this instance if we want to implement ray traced effects. There are zero new PSOs added in this case ...
This guide suggest using DXR 1.0 stile shaders instead of DXR 1.1 ubershaders.
It also suggest considering simplified shading and a unified shader for materials, which would result into fewer hit shaders.
This recommendation just sidesteps the issues of having many material models altogether. They're basically just asking developers to opt in to having fewer materials when ray tracing which isn't comparable at all to "uber-materials" ...

Yes, you can reduce the dispatch divergence/PSOs by removing other materials but you also see improvements when doing the same for inline RT as well such as reduced register pressure and shading divergence ...
 

OlegSH

Regular
Separate "RT shaders" (ray generation/intersection/miss/callable or closest/any-hit shaders) don't exist at all with inline RT. We add "RayQuery objects" to either existing compute or pixel shaders that already handle our material and lighting in this instance if we want to implement ray traced effects. There are zero new PSOs added in this case ...
There can be separate compute RT shaders for shadows, AO, and other stuff like that, why should they be necessary integrated into the main lighting pass shaders? Вut I wasn't talking about that. Inline code is still a code that has to be compiled. The number of PSOs depends not just on how much there are types of shaders, but rather on performance stats in runtime, compiling a few large PSOs with heavy ubershaders can cause noticeable frametimes spikes in comparison with compiling more PSOs with simpler shaders.

Yes, you can reduce the dispatch divergence/PSOs by removing other materials but you also see improvements when doing the same for inline RT as well such as reduced register pressure and shading divergence ...
The difference is that you can do simple and complex stuff with DXR 1.0 and you have to be quite creative to do the complex stuff with inlined RT due to performance and pipeline limitations (loops instead of recursion, etc).
 
So the final A770 specs are here, the A770 enjoys a healthy lead in memory bandwidth vs the RTX 3060 (560GB/s vs 360GB/s), operates at higher clocks (2100MHz vs 1800MHz), has access to a higher TDP (225w vs 170w), has higher count of FP32 cores (4096 vs 3584), and higher count of ray tracing cores (32 vs 28), so it's only natural the A770 beats the 3060 in Ray Tracing .. though, It's obvious it's beating it by raw brute force specs alone (35% higher FLOPs, and 55% higher bandwidth), and not by some better RT engine that Intel has. I think the Intel's RT engine is still slightly suboptimal to NVIDIA's engine, but it's a solid effort from Intel none the less.


 

Granath

Newcomer
So the final A770 specs are here, the A770 enjoys a healthy lead in memory bandwidth vs the RTX 3060 (560GB/s vs 360GB/s), operates at higher clocks (2100MHz vs 1800MHz), has access to a higher TDP (225w vs 170w), has higher count of FP32 cores (4096 vs 3584), and higher count of ray tracing cores (32 vs 28), so it's only natural the A770 beats the 3060 in Ray Tracing .. though, It's obvious it's beating it by raw brute force specs alone (35% higher FLOPs, and 55% higher bandwidth), and not by some better RT engine that Intel has. I think the Intel's RT engine is still slightly suboptimal to NVIDIA's engine, but it's a solid effort from Intel none the less.


even in this thread were tech presentation from Intel claiming that it's level 3.5, where for Nvidia is 3. RTFM please.
 

swaaye

Entirely Suboptimal
Legend
Yeah I was saying that too, but the 3000 series will still be around. They are still selling the 2060, 16x0 and some older low-end stuff after all. They are all competitive at the right price.
 
Last edited:
Top