• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

Will DLSS 3 add latency ?

Associate
Joined
22 Jun 2018
Posts
1,676
Location
Doon the watah ... Scotland
My understanding is that DLSS 3 will essentially interpolate between 2 ‘properly’ calculated frames and add in and display between them some interpolated frames … there by displaying more fps overall .. great idea.

But, in order to do that, it needs to calculate the later frame its going towards, then figure out the in between interpolated frames, display them, then finally display the calculated later frame. That will surely add latency into the display compared to simply displaying the later frame when it was originally calculated.

Admittedly, at high fps, its not much time between frames, but looking at the demo on the presentation, the original source was ~25fps (40ms/frame) boosted to 100+fps when you play at that wraps rate, there is a lot of inherit latency, and to wait for that latter frame to be reached via interpolated frames between is going to add more still surely, as in another 40ms ? So the display latency could be 80ms overall even though your seeing 100fps.

Or have I understood it all wrong?
 
i think the way it was described, dlss would take a higher resolution previous frame and lower resolution current frame and then transform the lower res current frame to its higher res counterpart with no loss in quality

its somewhat of an autocorrelation process that would be dynamically load balanced, there would be some way to intervene through feedback loops just so that the sequence doesnt degenerate into an undesirable state, so i dont believe latency could be introduced by low fps but there could be other ways it could get introduced in the pipeline
 
Last edited:
I would have to assume with the additional latency the overall latency still amounts to less than what you would get rendering native at a lower frame rate? Be kinda pointless to me otherwise...

Have to wait and see I guess :)
 
On the latency front, from nVidia's website - https://www.nvidia.com/en-us/geforce/news/dlss3-ai-powered-neural-graphics-innovations/

Scroll down a bit to where it talks about DLSS 3 with reflex - It shows that Cyberpunk 2077 is running at 4K Native at 23 FPS but has a hilariously odd 161ms latency but then when using DLSS 3 with Reflex it's now 101 but has 55ms latency.

They do know how sluggish 55ms feels, even on a single player game!?
 
My understanding is that DLSS 3 will essentially interpolate between 2 ‘properly’ calculated frames and add in and display between them some interpolated frames … there by displaying more fps overall .. great idea.

But, in order to do that, it needs to calculate the later frame its going towards, then figure out the in between interpolated frames, display them, then finally display the calculated later frame. That will surely add latency into the display compared to simply displaying the later frame when it was originally calculated.

Admittedly, at high fps, its not much time between frames, but looking at the demo on the presentation, the original source was ~25fps (40ms/frame) boosted to 100+fps when you play at that wraps rate, there is a lot of inherit latency, and to wait for that latter frame to be reached via interpolated frames between is going to add more still surely, as in another 40ms ? So the display latency could be 80ms overall even though your seeing 100fps.

Or have I understood it all wrong?

It calculates based on 2 previous frames to predict a new frame, not in between two existing frames, according to their description here.
And it suggests that if you are CPU limited the max frame boost is 2X - presumably due to latency and mismatch between what the user sees, the game engine has happening and its slow update, and the action the user then tries to do.

 
Last edited:
Frame generation will absolutely introduce a perception of latency, it generates a frame that doesn't come from the game AI/logic and therefore doesn't reflect the user input.

It's likely not a big deal for the likes of Flight Simulator.
 
It calculates based on 2 previous frames to predict a new frame, not in between two existing frames, according to their description here.
And it suggests that if you are CPU limited the max frame boost is 2X - presumably due to latency and mismatch between what the user sees, the game engine has happening and its slow update, and the action the user then tries to do.

Suggest you go back and read that again as that's not how it works. It interpolates between the current and previous frame to insert a new frame in-between the two. Therefore the current frame has to be delayed so the newly rendered intermediate can be drawn on screen first. This is why it operates in conjunction with reflex to minimise that latency hit.
 
Having a quick look about, articles suggest reflex reduces latency in good cases up to about 50%, and I assume normally less elsewhere. So it’s not going to fully address the doubling of latency created by waiting to display the current frame after the interpolated ones.

I suppose it’ll really depend on the game type though.
 
To answer your question yea dlss3 increases latency because it's adding in fake frames between real frames and each fake frame is a frame that doesn't accept player input, therefore increasing the latency between real frames that takeninput

However Nvidia has put Reflex into the pipeline to bring latency down - the question is, does reflex bring latency down enough or is latency still higher than without using dlss3? We won't know the answer until review day or even after as most reviewers will not test frame latency on their day 1 reviews - especially if there are no dlss3 games to test yet!
 
Last edited:
What will be interesting if it will always insert a fake frame in between real ones ( meaning dlss3 is locked to doubling the framerate) or if it will skip a few insertions, If the predictive system cannot reach the correct conclusion or for other reasons.

Skipping insertions would lead to frame rate yo-yoing which means latency will be highly inconsistent.
 
What will be interesting if it will always insert a fake frame in between real ones ( meaning dlss3 is locked to doubling the framerate) or if it will skip a few insertions, If the predictive system cannot reach the correct conclusion or for other reasons.

Skipping insertions would lead to frame rate yo-yoing which means latency will be highly inconsistent.

Yeah, it’s going to be interesting to see how it works. But I get the feeling it won’t be until next gen cards are out (whatever is after Ada) that it will get fixed just like DLSS 1.0 was crap. Will help sell a bunch of Ada cards though :cry:
 
I look forward to many many many many many many many many many many many many many many many many many many many many many many threads popping up with people zooming in 55X to spot differences with it on vs off.
 
I look forward to many many many many many many many many many many many many many many many many many many many many many many threads popping up with people zooming in 55X to spot differences with it on vs off.

To be fair, I think people will notice higher latency more than any difference on zoomed screenshots and if latency is any higher than just not using dlss3 then you can imagine no one will use it for latency sensitive games like esports
 
Last edited:
Back
Top Bottom