With the arrival of the new RTX 50-series line of graphics cards, Nvidia has released DLSS 4. While multi frame generation features are exclusive to the new GPUs, the existing DLSS super resolution and ray reconstruction technologies are upgraded with a far superior transformer model, which can be used on any RTX card, going all the way back to the 2018 Turing offerings. There are big improvements across the board and we’ll be looking at each component DLSS 4 offers, starting today with ray reconstruction. To what extent is quality improved? If the model is indeed more complex than the older convolutional neural network offering, what’s the performance impact on legacy RTX cards?
I started my investigation into DLSS 4 by looking at ray reconstruction because it’s a highly challenging technology – and thus far, Nvidia has no competition. The idea is simple: for ray tracing, hardware constraints mean we can’t trace unlimited rays. Therefore, there’s a lower resolution RT image that must be ‘denoised’ to produce something akin to a standard resolution image. But here’s the thing: if using DLSS upscaling, the denoised image is then upscaled to your output resolution – and RT effects will present poorly compared to the rest of the frame. Ray reconstruction denoises and ‘upscales’ RT to your desired resolution simultaneously – and while the first-gen effort was highly impressive, there were a number of issues and failure points.
This was especially obvious in Cyberpunk 2077 where you could find severe issues with smudging and ghosting on mid-distant moving objects or objects that were indirectly lit; you could find examples of oversharpening and ‘stylisation’ occurring, giving an oily and runny unstable look – and there could be issues with how art was displayed, particularly obvious in skin rendering. The good news is that the transformer model is a generational improvement in every respect, with only one or two outstanding issues I’d like to see resolved.
As is often the case for image quality comparison analysis pieces, video is far more illuminating than text, so check that out above if you have the means. However, kicking off with stylisation concerns, normal map detail is more accurately presented with the new tech, offering an immediate quality mode upgrade. With the CNN model, ray reconstruction even had problems resolving straight lines – not an issue now. The video showcases how the transformer model not only resolves more detail but does away with the CNN model’s ‘interpretation’ of the original art. It’s fundamentally different – and better – in Cyberpunk 2077 to the point where there’s a case that the original art was never properly represented in many respects until now.
The new ray reconstruction transformer model more properly respects other elements of game rendering, such as sub-surface scattering – an effect typically used to produce accurate-looking skin shading. With the old CNN version, it looked as though sub-surface scattering was invisible when ray reconstruction was active, giving people’s faces a papery and craggy look, almost ageing them. This is gone with the new model – less of the detail on the skin is visible with pin-point precision, while shadowed regions have a reddish glow to them now as light permeates through the skin’s surface.
The last – and arguably biggest – improvement with the transformer model comes from a heavy decrease in ghosting. Now, the CNN model had the propensity ghost in all games. Combine this with what I call stylisation artefacts along with the lack of sub-surface scattering and NPCs could look pretty poor. The difference is stark with the new DLSS: while some of the lighting detail is a little unstable, the cumulative effect of all the improvements is such that ray reconstruction in performance mode can even improve on the quality CNN mode, despite the huge gulf in native rendering resolution.
While the improvement is huge, there are still some issues I’d like to see addressed. For example, glass reflections look much the same. After the camera sits still for a while, both the old and the new models don’t know when to stop aggregating temporal data, so after sitting for a while without moving, objects in motion can smear and can look almost transparent in glass reflections and only moving the camera again makes them stop ghosting. Beyond that, there are two regressions I have found in my testing. One of them can be found in Alan Wake 2, specifically with the TV screens in the game. If the camera sits still, you can see the images being overly accumulated, leading to obvious smears and ghosts and other strangeness- much like the glass reflections. This does not happen with the old CNN model.
The other regression I have seen – at least in the performance mode – is a kind of vertical striation showing up in some parts of the image at times when the camera is still… but that’s about it. Further testing may show more issues to address, but right now, the aggregate improvement to quality is remarkable and obviously outweighs the negatives.
With that increased quality in mind, the question is the extent to which this new technology may not run as well on older RTX GPUs bearing in mind the higher computational cost. Looking at the RTX 5090 as a reference point, the cost of ray reconstruction using the transformer model is not a big deal – all the extra quality at the same input resolution as before, with a mere seven percent measured decreased average frame-rate over the course of the benchmark using ray tracing set to psycho at 4K quality mode.
| RTX 2080 Ti | RTX 3090 | RTX 4090 | RTX 5090 | |
|---|---|---|---|---|
| 4K Super Res CNN | 100.0% | 100.0% | 100.0% | 100.0% |
| 4K Super Res Transformer | 92.1% | 93.5% | 95.3% | 96.0% |
| 4K Ray Reconstruction CNN | 100.0% | 100.0% | 100.0% | 100.0% |
| 4K Ray Reconstruction Transformer | 64.7% | 68.7% | 95.2% | 93.0% |
On RTX 4090, I actually saw a better result, with the same settings seeing only a five percent decrease in performance with ray reconstruction in quality mode 4K while using the transformer model. Performance only starts becoming very important when going down to Ampere and Turing cards, based on my testing. On the RTX 3090, using the transformer model reduces average frame-rate by more than 31 percent over the course of the benchmark. On RTX 2080 Ti, it was even worse: a 35 percent reduction in average frame-rate over the course of the benchmark. So, it appears the transformer model for ray reconstruction is quite a lot heavier on Ampere and Turing GPUs. Unfortunately, reducing output resolution does not really seem to help. As the table above indicates though, thankfully the performance hit using the super resolution transformer model is nowhere near as impactful.
In summary, the transformer model has massively increased image quality in nearly all the core areas I tested. It has boosted it so much that in Cyberpunk 2077, if you put the transformer model in performance mode next to the balanced and quality mode of the CNN model, one can easily see many areas in the image where the lower resolution transformer image just looks better. There is far less ghosting, more coherent detail, and no stylisation.
Not every game and every aspect of image quality with the new transformer model will necessarily look better in performance mode than the CNN model in quality mode, but first impressions with Alan Wake 2 and Cyberpunk 2077 certainly suggest this is the case. As always though, more testing is going to be needed, so look forward to further coverage of DLSS 4.