Nvidia slip the show atCES 2025with theannouncement of the RTX 5090 , and despite mess of sermon about the card ’s $ 2,000 price shred , it ushers in a quite a little of new technology . Chief among them is DLSS 4 , which brings multi - frame generation to Nvidia ’s GPUs , offer a 4X performance boost in over 75 games right on away when Nvidia ’s newRTX 50 - seriesGPUs hit the street .
So , let ’s set the record flat , at least as much as I can before Nvidia ’s novel graphics card are here and we all experience what DLSS 4 has to offer first - hand .
No, it doesn’t ‘predict the future’
One of the independent issues about properly understanding how DLSS 4 kit and caboodle derive from a comment Nvidia CEO Jensen Huang made during a Q&A. Jarred Walton over atTom ’s Hardwareasked Huang about how DLSS 4 work on a expert spirit level , and Huang categorically deny that DLSS 4 use frame interpolation . He said that DLSS 4 “ predicts the future , ” instead of “ falsify the past . ” That ’s a buzzy quote , for sure . Too bad it ’s wrong .
DLSS 4 ’s multi - frame genesis uses a proficiency called framing interpolation . This is the same proficiency that we find out in DLSS 3 , and it ’s the same proficiency you ’ll determine in other frame generation toolslike Lossless ScalingandAMD ’s FSR 3 . Frame interpolation act like this : Your graphics card fork up two frames , and then an algorithm step in to calculate the dispute between those frames . Then , it “ give ” a chassis to go in between , gauge what the interstitial systema skeletale would look like base on the departure between the two frames that were rendered .
And DLSS 4 utilise frame interpolation . There has been some early research on new proficiency to generate frames — in particular , research fromIntel about systema skeletale extrapolation — but it ’s still other day for that technology . There ’s some point I ca n’t partake in quite yet , but I ’ve substantiate with multiple source now that DLSS 4 is , in fact , using frame interposition . It makes signified , too . These type of rendering tool do n’t just bug out up out of nowhere , and there ’s almost always a long lineage of enquiry papers before any new interpretation technique is turned into a marketable product like DLSS 4 .
That does n’t take out from what DLSS 4 is capable of . It might be using the same proficiency as DLSS 3 for creating new frames , but that should n’t perturb you from what DLSS 4 can actually do .
Latency isn’t the issue you think it is
I understand why Nvidia does n’t want to comment a good deal on DLSS 4 ’s exercise of frame interjection . That ’s because frame interpolation introduces response time . You need to return two frames and then do the interpellation before the first frame in the succession in displayed , so when using any frame insertion tool , you ’re essentially playing on a slight delay . The assumption I ’ve ascertain is that these extra frame linearly increase latency , which is n’t the case .
The Verge showed concernsaying it desire to “ see how the new skeletal system generation tech affects response time , ” whileTechSpot declaredthat “ users are concerned that multi - frame rendition could compound the [ latency ] problem . ” It ’s a natural buffet to the multiplied “ fake ” frames that DLSS 4 can spatter out . If beget one frame cause a latency job , surely beget three of them would cause a bigger reaction time job . But that ’s not how it knead .
This is why it ’s so important to understand that DLSS 4 uses frame interpolation . The estimation of playing on a postponement is n’t any different between DLSS 3 generating one extra human body and DLSS 4 generating three extra ones — the process still ask rendering two frame and comparing the difference between them . Your rotational latency does n’t importantly increase between inserting one , two , or three extra frames in between the two that were rendered . disregardless of the number of frame of reference that go in between , the rotational latency add by the frame interpolation process is largely the same .
Let me instance this . Let ’s say you ’re bring a game at 60 systema skeletale per 2d ( fps ) . That entail there ’s 16.6 milliseconds between each frame you see . With DLSS 3 , your skeletal frame rate would double over to 120 fps , but your latency is n’t halve to 8.3ms . The game looks smoother , but there ’s still 16.6ms between each try skeleton . With DLSS 4 , you ’ll be able to go up to 240 fps , quadruple your frame rate , but once again , the latency does n’t drop to 4.2ms . It ’s still that same 16.6ms .
This is a very reductive tone at PC reaction time — there ’s overhead for DLSS Frame Generation to persist , plus the latency add together by your monitor and mouse — but it ’s useful for understanding that the sum latency does n’t linearly increase when adding more frame to the frame interpolation process . The time between each rendered frame does n’t exchange . The latency you know is still mostly the result of your base frame rate before DLSS Frame Generation and the overhead that the instrument has .
You do n’t have to just take my intelligence for that . Digital Foundry has tested DLSS 4 , including latency , and found just what I just describe . “ It seems to me that the majority of the extra latency still come from soften that extra skeletal system , but adding further intermediate frames comes with a comparatively minimum increase in response time , ” wroteDigital Foundry ’s Richard Leadbetter . The humble amount of extra latency only arrive from DLSS calculate more frames in between the two that have been rendered , so the bulk of the rotational latency increase with DLSS 4 is n’t much unlike from DLSS 3 .
The rotational latency take with DLSS 4 is for the most part the same as it is with DLSS 3 . If you ’re playing at a low groundwork build charge per unit , there ’s a disconnection between the responsiveness you ’re experience and the smoothness you ’re visualise . That disconnect will be more significant with DLSS 4 , but that does n’t suddenly mean there ’s a monolithic growth in response time as a resultant role . That ’s why Nvidia ’s impressive Modern Reflex 2 is n’t require for DLSS 4 ; just like DLSS 3 , developers only take to go through the first version of Reflex for DLSS 4 to work .
A completely new model
clarify how DLSS 4 works might lead you to believe it ’s more of the same , but that ’s not the case . DLSS 4 is a very pregnant departure from DLSS 3 , and that ’s because it uses a whole different AI theoretical account . Or , I should say , AI models . As Nvidia details , DLSS 4 run five separate AI models for each rendered frame when using Super Resolution , Ray Reconstruction , and Multi - Frame Generation , all of which need to carry out in a issue of milliseconds .
Because of what DLSS 4 entails , Nvidia did away with its previous Convolution Neural web , or CNN , and it ’s now using a vision transformer mannikin . There are two large changes with a transformer model . First is something called “ self - attention . ” The mannequin can give chase the importance of different pixels over multiple frames . Being self - referential in this elbow room should allow the new model to focus more on problematic surface area , such as thin detail with Super Resolution that may show shimmering .
Transformer example are also more scalable , allowing Nvidia to add far more parameters to DLSS than with the late CNN approach . accord to the company , the raw transformer model has double the parameters , in fact .
I ’ve seen DLSS 4 in action a couple of clip , but the real test for the feature article will be when Nvidia ’s next - gen GPUs launching . Then , I ’ll be able-bodied to evaluate how the feature of speech works across several games and scenario to see how it holds up . Regardless , there are a lot of change with the characteristic , and according to what Nvidia has shared so far , those changes work to make DLSS even well .
