DLSS 5 is coming a lot sooner than you might think. You've probably seen the leaked DLLs and the various mods putting the technology into existing games, but DLSS 5 will officially arrive later this week, on September 3, launching in exactly one title: NBA 2K27.
That's right, the launch title for DLSS 5 is a basketball game.
We have some interesting information to share today, including Nvidia's first-party performance data for the official implementation of DLSS 5, although not quite as much as we would have liked. A proper hands-on analysis will have to wait until we can test it ourselves in NBA 2K27.
Nvidia's DLSS 5 Performance Claims
Let's get into the biggest question we had about DLSS 5: performance. Not the performance of the feature when modded into games using RenoDX or OptiScaler, but the actual performance of an official game implementation. Nvidia provided some rather... interesting performance charts showing exactly that.
At 4K, Nvidia is showing NBA 2K27 running with ray tracing enabled and the Ultra preset. It's a slightly strange chart because only two GPUs are included: the RTX 5080 and Nvidia's flagship RTX 5090. And boy, are those frame rates high.
Then you check the footnotes.
The results were captured using Multi Frame Generation in 6X mode, with DLSS Super Resolution set to Performance mode. When you strip away 6X frame generation, the RTX 5080's 233 FPS result implies a render rate of just 39 FPS. Multi Frame Generation then turns that into the final output rate shown on the chart. The RTX 5090's 370 FPS result similarly implies a render rate of about 62 FPS with DLSS 5 enabled.
That also helps explain why Nvidia only shows those two GPUs at 4K. Anything below the RTX 5080 would presumably be operating at an even lower render rate, despite potentially producing a very high final frame rate after 6X Multi Frame Generation.
Actual render performance should be somewhat higher with frame generation disabled because frame generation itself carries an overhead. Nevertheless, the RTX 5080 will almost certainly run DLSS 5 with Performance upscaling at a render rate below 60 FPS at 4K.
At 1440p, Nvidia provided more data, this time using Quality upscaling.
The RTX 5090 reaches 594 FPS with 6X Multi Frame Generation, implying a render rate just shy of 100 FPS, which is perfectly usable. The RTX 5080 drops to an implied 69 FPS, while the RTX 5070 Ti sits at around 59 FPS and the RTX 5070 at roughly 44 FPS.
Realistically, based on those figures, it doesn't look like the RTX 5070 will be able to use DLSS 5 at a particularly attractive render rate at 1440p. The RTX 5070 Ti looks borderline depending on how comfortable you are with a roughly 60 FPS base rate before frame generation.
At 1080p, the doors open to more GPUs. The RTX 5090 has an implied render rate of roughly 133 FPS, the RTX 5080 around 100 FPS, the RTX 5070 Ti about 88 FPS, the RTX 5070 around 64 FPS, the RTX 5060 Ti at 54 FPS, and finally the RTX 5060 at just 43 FPS.
We say the doors open to more GPUs, but we're still talking about an RTX 5070 rendering at barely over 60 FPS at 1080p with Quality upscaling enabled, while cards below that fall into sub-60 FPS territory. Ouch.
The Ugly Catch Behind Those Huge Frame Rates
The big kicker for DLSS 5 performance comes from another Nvidia slide. It shows NBA 2K27 with DLSS disabled running at 227 FPS, while enabling DLSS increases the displayed frame rate to 370 FPS. Except, of course, those are very different rendering configurations.
The 227 FPS result on the left is native 4K rendering. On the right, DLSS Super Resolution is set to Performance mode, Multi Frame Generation is running at 6X, and DLSS 5 Neural Rendering is enabled. That 370 FPS output again implies a render rate of roughly 62 FPS. In other words, the underlying render rate appears to fall from 227 FPS in the native example to around 62 FPS in the DLSS 5 configuration, despite the latter also using Performance upscaling.
These are first-party benchmarks, and Nvidia has not provided an isolated DLSS 5 on vs DLSS 5 off performance comparison using otherwise identical settings, so we'll need independent testing before drawing firm conclusions about the exact cost of Neural Rendering itself.
Nvidia has not provided an isolated DLSS 5 on vs DLSS 5 off performance comparison.
But based on what Nvidia is presenting, DLSS 5 appears to be an extremely taxing feature. And that's despite Nvidia saying it has made significant strides in optimizing the model since its first reveal in March.
Does DLSS 5 Actually Look Better?
That naturally raises another question: are the visual gains worth the performance cost? It's the same basic question that surrounded ray tracing when it first arrived (and to this day).
There are also plenty of questions about how it holds up in motion, particularly because DLSS 5 operates on the rendered frame. Nvidia says the technology processes frames individually while using engine data such as motion vectors to maintain temporal stability, rather than operating on a sequence of frames like a conventional video-generation model.
That makes motion stability particularly interesting to investigate. DLSS 5 is modifying lighting and material responses in the final image while working from the information available for the current rendered frame, so consistency from one frame to the next will be crucial.
Nvidia provided several NBA 2K27 comparisons showing DLSS 5 enabled and disabled. Unlike the various mods we've seen in the last week, this is the final implementation built directly into the game, with Visual Concepts controlling how the technology is applied.
In other words, this is how the developers have chosen to use DLSS 5. It's part of their intended visual presentation.
According to Nvidia's figures, the "before" example can run at more than 220 FPS natively, while the DLSS 5 configuration has an implied render rate of around 62 FPS before being interpolated to 370 FPS using Multi Frame Generation.
Again, I don't want to make too many judgments about image quality from still shots. If DLSS 5 doesn't hold up in motion, it will be essentially worthless regardless of how impressive individual screenshots might look.
Nvidia Has Seriously Toned Down DLSS 5
But one observation I do have from this announcement is that Nvidia has drastically toned down DLSS 5's output compared with the original reveal in March. And I mean seriously drastically.
The over-the-top examples that drew accusations of looking like AI slop are nowhere to be found in this latest presentation. They've been replaced by much subtler examples that are ultimately less transformative.
Maybe that's a change for the better: less of a slopified look, at the expense of DLSS 5 "doing less," for lack of a better term.
Pushing for "Artistic Intent"
Nvidia has also changed its messaging around DLSS 5 to hammer home the idea that the technology is designed to preserve artistic intent. This has become a hugely contentious topic around DLSS 5, and the issue came up again recently when the leaked version was modded into existing games.
In many demonstrations, users simply cranked the feature to maximum intensity, dramatically changing the appearance of games with zero input from their original developers.
Nvidia repeatedly emphasizes that DLSS 5 is "trained to preserve artistic intent", that "artists direct the final frame" and that DLSS 5 "respects the rendered frame". It also enables "artist-directed output" with control over the final output. I'm not joking here, about half of this announcement is related to this topic.
Developers are given controls such as Structure Intensity and Tone Intensity, along with model selection and masking tools that can change how DLSS 5 affects individual objects or areas of a scene. Those controls aren't intended to be exposed to gamers through normal graphics settings. For players, Nvidia's official implementation is essentially an on-or-off option.
Therefore, in my opinion, if developers choose to integrate DLSS 5, then the DLSS 5 output becomes part of their artistic intent, just like the output produced by every other graphics setting in the game. Mods and demonstrations created without developer involvement are obviously different.
You can install a DLSS 5 mod and absolutely destroy a game's artistic vision if you want, but that's not fundamentally different from installing ReShade and badly altering the appearance of a title.
The bigger issue with DLSS 5's visual output isn't whether it follows the developer's artistic vision. It's whether the result actually looks good.
And that's where DLSS 5 is going to become very interesting. Whether its output looks better is ultimately a subjective question, and given the reaction to DLSS 5 so far, I suspect opinions will vary enormously between gamers, probably more than with almost any graphics feature we've seen before. Even among Nvidia's latest examples, there are some I prefer to others, and I'm sure my choices won't line up with everyone else's.
That brings us straight back to performance. If DLSS 5 really is as computationally expensive as Nvidia's first-party numbers suggest, it will be difficult to convince gamers to enable it unless the visual improvement is both significant and attractive.
What Does DLSS 5 Actually Do?
The main goal of DLSS 5 is to produce more realistic lighting and material responses. It does this through what Nvidia calls 3D-Guided Neural Rendering, using a pixel-space diffusion transformer model. The model targets lighting interactions and material properties, analyzing information from the game engine and processing the rendered frame in an attempt to produce more realistic results.
Skin is one obvious example. DLSS 5 attempts to generate more realistic interactions between light and skin, which involves complex effects such as subsurface scattering. But there are also plenty of things DLSS 5 can't improve, and Nvidia provides a good example of that.
Across Nvidia's examples, you'll see changes to the intensity of lighting effects, tone mapping, material responses, and how surfaces interact with light.
There are a number of low-quality textures in one demonstration scene, and DLSS 5 does nothing to transform them into higher-resolution assets. If the geometry of an object is extremely basic, it remains basic after DLSS 5 is applied. It also can't simply turn a low-quality reflection into a detailed ray-traced reflection. In Nvidia's example, the resolution and underlying structure of a ground reflection remain largely the same in both versions.
It's a similar story with existing shadows. DLSS 5 isn't increasing the resolution of a low-resolution shadow map that has already been rendered by the game. What it can do is alter the perceived shadowing of the scene by adding effects such as deeper contact shadows and ambient-occlusion-like shading based on what the model determines is appropriate for the scene.
Importantly, unlike ray tracing, these additional lighting effects aren't physically simulated rays. They're generated by Nvidia's neural rendering model based on the scene information provided to it and the data used to train the model. Because DLSS 5 can't replace or reconstruct every element of a scene, Nvidia says better input imagery produces better output.
Nvidia demonstrates this using rasterized and ray-traced inputs. The final DLSS 5 result differs depending on the underlying rendered image because the model is building on the lighting and scene information it receives rather than completely replacing the game's rendering pipeline.
This makes us particularly interested in what DLSS 5 will do in games that already use cutting-edge rendering techniques, especially titles with high-quality physically based materials, ray tracing, or full path tracing.
The "before" shots Nvidia has shown from NBA 2K27, for example, aren't necessarily what we'd describe as the absolute cutting edge of modern real-time rendering.
Can Older GeForce GPUs Run DLSS 5?
For the official launch, Nvidia has only announced DLSS 5 support for the GeForce RTX 50 series. The company hasn't announced support for RTX 40-series or older GPUs. But Nvidia has now been put in an interesting position by the leaked DLL. The initial build was restricted to RTX 50-series GPUs, but modders quickly found ways to get the technology running on RTX 40-series hardware.
Modders have even managed to get DLSS 5 running on RTX 20- and 30-series GPUs, initially with extremely poor performance, although the situation has improved as the unofficial implementations have developed (in one week!). Whether Nvidia ever chooses to officially support those architectures is another matter.
Performance may make the question largely academic for some of them anyway. Nvidia's own numbers suggest an RTX 5070 is already operating at relatively low render rates with DLSS 5 enabled, and that GPU is broadly in the performance class of older high-end cards such as the RTX 3090.
Still, if the technology can run on older hardware, there's an argument for allowing users to make that choice themselves.
As mentioned earlier, we haven't had much opportunity to explore DLSS 5 firsthand and to give any meaningful assessment of DLSS 5's image quality, we need to see how it behaves in motion, whether lighting remains consistent from frame to frame, how frequently artifacts appear, and whether the visual improvements justify such a substantial performance cost.
We'll be digging into all of that once NBA 2K27 and the official DLSS 5 implementation become available.