Nvidia’s DLSS 4 is not what you think it is. Let’s dispel the myths

Nvidia stole the present at CES 2025 with the announcement of the RTX 5090, and regardless of a lot dialogue concerning the card’s $2,000 price ticket, it brings a variety of new know-how. Chief amongst them is DLSS 4, which brings multi-frame technology to Nvidia’s GPUs, delivering a 4X efficiency enhance in over 75 video games as quickly as Nvidia’s new RTX 50 collection GPUs hit the streets .

Nevertheless, I’ve seen a variety of misconceptions about how DLSS 4 truly works. Between deceptive feedback from Nvidia’s CEO and a radical redesign of how DLSS works, it is no marvel there’s been misinformation circulating concerning the new know-how, what it is able to to do and, critically, what limitations there are.

So let’s set the file straight, at the very least as a lot as I can earlier than Nvidia’s new graphics playing cards are right here and all of us expertise what DLSS 4 has to supply first hand.

No, it does not ‘predict the longer term’

One of many essential points about correctly understanding how DLSS 4 works comes from a remark that Nvidia CEO Jensen Huang made throughout a Q&A. Jarred Walton at Tom’s {Hardware} requested Huang how DLSS 4 works on a technical stage, and Huang flatly denied that DLSS 4 makes use of body interpolation. He mentioned DLSS 4 “anticipates the longer term”, quite than “interfering with the previous”. That is a buzz quote, to make certain. Too unhealthy it is incorrect.

Huang has waxed poetic about DLSS body technology previously, and whereas this type of framing works to clarify a know-how like DLSS 4 to a lay viewers, it additionally results in some misunderstandings about the way it truly works. After this quote, I truly had a number of readers inform me that I used to be misunderstanding how DLSS 4 works. I am not misunderstanding the way it works, apparently, however I perceive why there may be a variety of confusion.

DLSS 4’s multi-frame technology makes use of a way known as body interpolation. This is similar method we noticed in DLSS 3, and it is the identical method you may discover in different body technology instruments like Lossless Scaling and AMD’s FSR 3. Body interpolation works like this: Your graphics card renders two frames, after which an algorithm steps in to calculate the distinction between these frames. It then “generates” a body to go in between, guessing what the interstitial body would seem like based mostly on the distinction between the 2 frames that got.

And DLSS 4 makes use of body interpolation. There was some early analysis into new strategies to generate frames – specifically, analysis from Intel about body extrapolation – however it’s nonetheless early days for that know-how. There are some particulars I can not absolutely share but, however I’ve confirmed with a number of sources now that DLSS 4 is, in reality, utilizing body interpolation. It is smart too. These sorts of rendering instruments do not simply seem out of nowhere, and there is nearly at all times an extended line of analysis earlier than any new rendering method turns right into a marketable product like DLSS 4.

This does not take away from what DLSS 4 is able to. It might be utilizing the identical method as DLSS 3 for creating new frames, however that should not take away from what DLSS 4 can do in it actually.

Delay shouldn’t be the problem you suppose it’s

I perceive why Nvidia does not need to remark an excessive amount of on the usage of body interpolation by DLSS 4. It’s because body interpolation introduces lag. You need to render two frames after which do the interpolation earlier than the primary body within the displayed sequence, so while you use any body interpolation instrument, you are basically taking part in with a small delay. The belief I’ve seen is that these additional frames linearly enhance latency, which isn’t the case.

The Verge expressed concern saying it wished to “see how next-generation framerate know-how impacts latency,” whereas TechSpot acknowledged that “customers are involved that multi-frame rendering may compound the (latency) downside. It is a counterpoint to inherent to the a number of “dummy” frames that DLSS 4 can throw out. If producing one body causes a latency downside, producing three of them would possible trigger an even bigger latent downside. Nevertheless it does not work like that.

For this reason it’s so vital to grasp that DLSS 4 makes use of body interpolation. The thought of ​​taking part in with a delay isn’t any totally different between DLSS 3 producing one additional body and DLSS 4 producing three additional frames – the method nonetheless entails rendering two frames and evaluating the distinction between them. Your delay does not enhance considerably between inserting one, two or three additional frames between the 2 given ones. Whatever the variety of frames moving into between, the delay added by the body interpolation course of is basically the identical.

Let me illustrate this. To illustrate you are taking part in a sport at 60 frames per second (fps). This implies there are 16.6 milliseconds between every body you see. With DLSS 3, your body charge would double to 120 fps, however your latency is not halved at 8.3 ms. The sport seems smoother, however there are nonetheless 16.6ms between every rendered body. With DLSS 4, you can go as much as 240 fps, quadrupling the body charge, however as soon as once more the lag does not drop to 4.2 ms. It’s nonetheless the identical 16.6 ms.

This can be a very reductive take a look at PC latency – there’s overhead for producing DLSS frames, plus the added latency from the monitor and mouse – however it’s helpful to grasp that base latency does not enhance linearly as extra are added many body to border interpolation course of. The time between every given body doesn’t change. The lag you expertise continues to be largely a results of your base framerate earlier than producing DLSS frames and the overhead the instrument has.

You do not have to take my phrase for it. Digital Foundry has examined DLSS 4, together with latency, and located precisely what I simply described. “It appears to me that many of the further latency comes from saving the additional body, however including different intermediate frames comes with a comparatively minimal enhance in latency,” wrote Digital Foundry’s Richard Leadbetter. The small quantity of further delay comes merely from DLSS calculating extra frames between the 2 given, so many of the delay enhance with DLSS 4 is not a lot totally different from DLSS 3.

The lag downside with DLSS 4 is basically the identical as with DLSS 3. When you’re taking part in at a low base framerate, there is a disconnect between the responsiveness you are experiencing and the smoothness you are seeing. This disconnect will probably be extra important with DLSS 4, however that does not all of the sudden imply there is a large enhance in latency in consequence. That is why Nvidia’s spectacular new Reflex 2 is not required for DLSS 4; like DLSS 3, builders solely have to implement the primary model of Reflex for DLSS 4 to work.

A very new mannequin

Explaining how DLSS 4 works would possibly lead you to imagine that it is extra of the identical, however it’s not. DLSS 4 is a really important departure from DLSS 3, and that is as a result of it makes use of a totally totally different AI mannequin. Or, I ought to say, AI fashions. As Nvidia particulars, DLSS 4 runs 5 separate AI fashions for every body rendered when utilizing Tremendous Decision, Beam Reconstruction, and Multi-Body Technology, all of which ought to run inside milliseconds.

Due to what DLSS 4 entails, Nvidia ditched its earlier Convolution Neural Community, or CNN, and is now utilizing a imaginative and prescient transformer mannequin. There are two main variations with a transformer mannequin. First is one thing known as “self-attention.” The mannequin can monitor the significance of various pixels in a number of frames. Being self-referential on this means ought to enable the brand new mannequin to focus extra on downside areas equivalent to positive particulars with Tremendous Decision that may present glare.

The transformer fashions are additionally extra scalable, permitting Nvidia so as to add many extra parameters to DLSS than with the earlier CNN strategy. In response to the corporate, the brand new transformer mannequin has double parameters, in reality.

As you may see within the movies above, Nvidia claims that this new mannequin has higher stability and preservation of positive particulars in comparison with the earlier CNN strategy. These enhancements aren’t unique to RTX 50 collection GPUs both. All RTX graphics playing cards will be capable of use the brand new transformer mannequin in DLSS 4 video games, at the very least for options supported by every technology.

I’ve seen DLSS 4 in motion a number of instances, however the actual check for the characteristic will probably be when Nvidia’s next-gen GPUs launch. Subsequent, I will be capable of consider how the characteristic works in a number of video games and eventualities to see the way it works. Regardless, there are many adjustments to the characteristic, and in line with what Nvidia has shared up to now, these adjustments work to make DLSS even higher.