Hands-on with DLSS 4 on Nvidia’s new GeForce RTX 5080

Last night, Nvidia finally unveiled the RTX 50-generation graphics cards – but over the holiday period, I had the opportunity to spend some time testing out the brand new DLSS 4 upscaling and frame generation technology on a PC kitted out with the new GeForce RTX 5080, running an updated build of Cyberpunk 2077 – and it’s impressive. Nvidia has upgraded its super resolution upscaling and ray reconstruction technology with a new ‘vision transformer’ model, offering some dramatic quality upgrades, while frame generation is boosted from one interpolated frame to two – or even three. The end result is a better-looking Cyberpunk 2077 capable of running the full path-traced experience at frame-rates well, well north of 120fps.

There’s a lot to discuss here but to be clear, this is preview, first look coverage. The RTX 5080 I had access to is an engineering sample. The drivers are not final. I can offer a broad idea of how DLSS 4 works and ballpark frame-rate increases, but exact numbers will have to wait for review hardware and final drivers. Also, the limited time window I had with the hardware limits the extent of the testing I could carry out – but regardless, I saw enough and captured enough to have an initial response to the new technology.

The fruits of my visit are helpfully embedded below, but speaking of capture, just like DLSS 3 frame-gen before it, showing off how DLSS 4 actually looks is somewhat difficult, owing to the frame-rate amplification factor of the new frame generation system. Multi-frame generation makes DLSS 4 a good fit for the latest wave of QD-OLED 4K 240Hz monitors – but there is no capture technology on the market capable of capturing ultra HD at 240 frames per second.

In the video b-roll, I tried to show the quality of frame generation by capping to 120fps, then slowing down by 50 percent to fit within the limits of YouTube’s top-end 4K 60fps container. However, running unlocked, frame-rates are much higher than the 120fps limit I imposed. So while the video gives you an idea of how the new frame-gen presents, the real-world experience is something quite different: frame persistence will be significantly lower and therefore frame generation artefacts are far less noticeable. In effect, I needed to hold back DLSS 4 to give any kind of representative media on a video platform. DLSS 4 with full multi-frame-gen enabled is designed for the latest generation of high refresh rate displays and I experienced it on an Alienware AW3225QF QD-OLED 4K 240Hz monitor running fully unlocked – and it’s quite the thing.

Thankfully, improvements to DLSS super resolution and ray reconstruction can be shown with 4K 60fps capture – and there are key improvements here. The existing DLSS uses a convolutional neural network – Nvidia says this “generates new pixels by analysing localised context and tracking changes in those regions over successive frames”. The model has improved over time, but it can only go so far and the latest releases have shown only iterative improvements.

The new DLSS model uses a vision transformer, similar to the base technology behind the likes of ChatGPT, Gemini and Flux. Nvidia says this “enables self-attention operations to evaluate the relative importance of each pixel across the entire frame and over multiple frames.” There are twice as many parameters as the prior CNN model and as a result, we’re promised greater stability, reduced ghosting, higher detail and enhanced anti-aliasing. The transformer model is highly scalable, so Nvidia is predicting further improvements to quality with better training. It’s also backwards compatible, meaning it can be retrofitted to prior DLSS titles running on the 2.0 version or higher. It’ll also work on all prior RTX cards, going back to 2018’s Turing-based RTX 20-series but the increased complexity of the model may have some kind of performance impact (possibly offset by the increase to quality – this really needs thorough testing).

Even though the captured material we had was limited, the general impression that image quality specialist Alex Battaglia and I have from the b-roll is that despite running at the same 1080p internal resolution as the CNN performance mode, the vision transformer equivalent doesn’t look like it – it’s better. And there are some clear signs that some of the long-standing issues we’ve had with DLSS are much improved. This is best shown in the video above, but smearing and ghosting are drastically reduced. Shimmer in motion and unwanted ‘stylisation’ of content are mostly gone, Meanwhile, ray reconstruction – essentially upscaling for ray tracing – shows similar, highly welcome improvements.

To compare the vision transformer model to the existing CNN version of DLSS is essentially to compare two very different AI upscalers – well outside of the scope of the hours I had with the RTX 5080 and something we’ll be looking at once we’ve got hardware in-house. However, the implications of the super resolution and ray reconstruction improvements are significant. DLSS has enjoyed only iterative improvements since the arrival of the 2.0 version in 2020 – and yet it remains the highest quality upscaler around, while DLSS ray reconstruction has yet to be delivered by any of the competition. Assuming dramatic improvements with no downside are delivered on both counts, this takes Nvidia a further step ahead of its rivals.

DLSS frame generation also gets new features – but these are tied exclusively to the new RTX 50-series cards. The single generated frame-gen tech used in 40-series is augmented with two and three frame generation in 50-series. This is tricky stuff, particularly when it comes to pacing those intermediate frames in a smooth, consistent manner – so Nvidia says that the new Blackwell architecture contains a hardware component for ensuring even pacing.

I used the black market in the Petrochem stadium for testing – an area we’ve previously used for identifying a graphics bottleneck in the console version of Cyberpunk 2077. Running fully unlocked at 4K output resolution with the RT Overdrive path-tracing setting enabled, the existing single frame generation tech combined with the vision transformer model for super resolution and ray reconstruction saw a 535 percent frame-rate multiplier compared to the game running at native 4K resolution. The gains continue as we made more generated frames. Adding two frames sees that rise to 725 percent, while three frames predictably delivers the biggest increase of all – 913 percent vs native resolution rendering.

Comparing the RTX 5080 with full frame generation and the transformer model enabled to the outgoing 4080 Super with single frame-gen and the existing CNN super resolution/ray reconstruction tech, I logged an improvement of 91 percent to frame-rate.

Returning to frame-pacing, you’ll see some visualisations of frame-times comparing single frame generation to full, three frame-gen output, both running on RTX 5080 – and it is reassuring to see that the latter not only means faster frame-rates, but also smoother frame-rates too. I took metrics here using a new version of Nvidia FrameView, but verified them with FCAT – where our software measures frame persistence based on the output of the GPU. Both results tallied.

My concern with the new frame generation technique was latency. When DLSS 3 launched, frame generation was achieved by buffering an extra frame then calculating the intermediate one. Both of these factors add latency, which was offset to a degree by the mandatory inclusion of Nvidia Reflex to claw back precious milliseconds. In the video, you’ll see that I’ve included PC latency results – again measured by FrameView. The averages across a circa two and a half minute run through the black market work out like this: a 50.97ms average latency with a single generated frame, rising to 55.5ms with two generated frames, and 57.3ms with the full, three-frame option.

Based on this sample then, we’re looking at a 6.4ms average addition to latency with maxed out frame generation in exchange for a 71 percent increase in frame-rate – it’s an equitable trade overall. Looking at these results, it seems to me that the majority of the extra latency still comes from buffering that extra frame, but adding further intermediate frames comes with a relatively minimal increase in latency. That means that the gameplay is still responsive in Cyberpunk and unless you’re super-attuned to input lag, you’re unlikely to tell the difference with the existing DLSS frame generation solution.

My time spent with the RTX 5080 was promising then. It’s still quite remarkable to see a path-traced Cyberpunk 2077 running well in excess of 120 frames per second with the new frame generation option, so it’s likely that less demanding games will run a whole lot faster. And then there’s the fact that frame generation has proven quite useful in overcoming CPU limitations in many games – it’ll be interesting to see the role multi frame-gen plays there. Meanwhile, the super resolution and ray reconstruction quality improvements look to address some of the key concerns we’ve had with DLSS, despite its status as the current highest quality upscaling solution.

We’ll be testing all components of DLSS 4 in much, much more depth as soon as we can, and – of course – we’ll be reviewing all of the RTX 50-series GPUs in due course.