RT + DLSS concern

[WereCat]

I dont own a 2000 series card and I likely ever wont have one but I'm still interested in the tech but I haven't been much active on this forum lately so I may have missed most of the discussion but I have been following the 2000 series launch pretty closely and I never seen anyone express the concern I have about running DLSS + RT at the same time. 

Since there is no game out that supports both, right now we can only speculate. 

 

My concern is that RTX card uses Tensor Cores for DLSS which is basically using an already trained neural network to take a frame on a given resolution and work it into a higher resolution which should make the image sharper so its like an AA except its supposed to have a performance benefit in terms of FPS vs a traditional AA. 

 

OC3D mentioned that when testing DLSS in the Final Fantasy benchmark (which is only available or was only available for the press at the time) he got more FPS than with no DLSS. Now I dont remember exacly whether that comparison applies in terms DLSS vs traditional AA or DLSS vs no AA at all but either way, there was a FPS uplift. 

 

Now RayTracing is supposed to use the same Tensor Core portion of the GPU die as DLSS but for denoising the RayTraced image. 

 

There goes my concern. 

Both DLSS and RT will use Tensor Cores for two things that have to happen in real time on multiple frames per second. Do these cards even have Tensor Core performance capable of doing both tasks at the same time without an either severe performance hit or missed frames on which they are not applied? 

 

I dont remember ever seeing a demo where they have ran both at the same time. 

[Tellos]

@WereCat I don't think it's a matter of them doing it at once, so much as having enough to do both split with half the tensor cores available. eg: half do DLSS and other does Real time ray tracing. Also the RTX chip does the live ray tracing as I understand it not the tensor cores. Though possible the Tensor cores play a role.

[WereCat]

2 minutes ago, Tellos said:

@WereCat I don't think it's a matter of them doing it at once, so much as having enough to do both split with half the tensor cores available. eg: half do DLSS and other does Real time ray tracing. Also the RTX chip does the live ray tracing as I understand it not the tensor cores. Though possible the Tensor cores play a role.

RT cores do the Ray Tracing. 

But that will leave a lot of noise on the image since you are limited with the amount of rays you can trace so they use the neural network for denoising which uses Tensor Cores. 

So now the card has to use Tensor Cores for denoising AND for enhancing the image resolution with DLSS. 

[Tellos]

@WereCat Then whats the RTX chips job? Seems odd.

[WereCat]

5 minutes ago, Tellos said:

@WereCat Then whats the RTX chips job? Seems odd.

The chip has CUDA cores, Ray Tracing cores and Tensor Cores.

CUDA cores are used all the time because thats what is standard right now. 

Ray Tracing cores are used only when you do Ray Tracing. 

Tensor Cores are used whenever you do anything with neural network which means that its at the very least intended for denoising the RayTraced image and DLSS. 

[pas008]

might be like a few gens away for both lol

[Guest]

i havent looked into it but you seem to have some valid points?

 

i have another concern for dlss alone. neural networks arent perfect and that could mean the image quality could be worse than other methods - we shouldnt just look at fps, we need to compare the images in reviews (just like way back)

[WereCat]

1 hour ago, Neftex said:

i havent looked into it but you seem to have some valid points?

 

i have another concern for dlss alone. neural networks arent perfect and that could mean the image quality could be worse than other methods - we shouldnt just look at fps, we need to compare the images in reviews (just like way back)

The enhancement from neural network done with Tensor Cores definitely is not accurate but the point there is that you get close to or indistinguishable accuracy to the real thing based on constant repetitive training which will eventually remove most of the noticeable errors by repetition alone.

For example, look at Deep Fakes.

[xg32]

4 hours ago, WereCat said:

.

I'll start off by saying i only bought the card for 4k gaming, not RT or DLSS. I think neither are really ready, just wait for next gen.

 

I believe DLSS will be useful, it's a cheap upscaling method that will improve regular and RT performance, regardless of what the resouce allocation is.

 

There's been report of devs toning down ray tracing to get 1080p/60fps, and the metro demo didn't even look at that impressive. I'm not really concerned about the tensor cores overworking themselves with both on, there will be a final number that comes out, and if it's garbage, turn RT off

 

Again, if you are looking to RT, skip this gen. The windows october update will hopefully let us RT our desktops