A look at Nvidia DXR, Broadcast App and DLSS visual quality

Ray-tracing reflections in Cyberpunk and DLSS 2.0

We are introducing a new testing methodology for graphics cards. One of the reviewed cards will be the GeForce RTX 3070 and we have an interesting one: Aorus GeForce RTX 3070 Master 8G with six video outputs. Before the review comes out, we’ve use it for a visual comparison to see the impact of ray tracing and how the different DLSS settings in the latest Cyberpunk 2077 patch will affect the quality of the game graphics.

Now let’s move to the promised look at how Cyberpunk 2077 gets rendered with the latest patch 1.1. We were interested in the visual impact of raytracing effects and DLSS. For comparison of reflections without ray tracing and with it, we first have a few images without DLSS. We recommend that you click on the images and switch between them in your browser and. Differences that you would not perceive easily with the two images sitting next to each other on one screen stand out much more when you flip back-forth between two overlayed images.

Most of the concrete surface (floor) seem to not look visibly “better” with ray tracing, but that is likely because we as spectators do not know exactly what the creator of the scene intended to show is, exactly. But notice the corner with the paper or cardboard strip lying at the top edge of the cutout. With DXR turned on, there is a noticeable shadow just as could realistically be expected, while without DXR, that shadow is missing.

Furthermore, it can also be seen that without DXR, only a rough simulation of reflection on the water and wet surface is used. The machines appear in the puddle only as a smudge and also the color tint from the rid lighting is applied only in this zone. When reflections are handled by DXR, you can see that the red light affects a much larger part of the scene, including  concrete that is less water-covered outside the puddle zone. This should be more in line with how it would look reality. But the biggest difference is clearly what the reflection of the machine looks like: with DXR, its geometry and shape are shown isntead of a rough smudge and we can even see the reflection of the building on the right.

Sharpness of DLSS 2.0

In this puddle scene, DLSS 2.0 seems to be surprisingly good at hiding the amount of upscaling that is taking place – this may be because the scene it’s static, so temporal filtering is better able to “reconstruct” the image from successive frames. If the background and objects were moving, this effect would probably not work as well anymore and the lines and details would turn out to be coarser in faster modes (Performance, Ultra Performance). It would probably be more obvious that they are being upscaled.

DLSS in Quality mode should upscale the image from two-thirds resolution (66%), but note that this means 66% in one axis, not 66% of pixel count. Instead of 3840 × 2160 pixels, the image is rendered at 2560 × 1440. The balanced mode reduces the factor to 56% in both axes. The Performance mode then has a factor of 50%, so it performs exactly 2× scaling in both axes, so the number of source pixels is only a quarter (for 4K output, only 1920 × 1080 is rendered). DLSS in Ultra Performance mode is even more extreme, with a factor of 33%. This means that 3840 × 2160 is obtained by 3x magnification in both axes, so the actual rendering resolution is only 1280 × 720 pixels (one-ninth of total pixels…). This mode is intended mainly for 8K output – in that case, the real rendering resolution is 2560 × 1440.

You can spot differences between Quality and Ultra Performance. For example, upscaling and temporal filtering to a higher resolution caused a loss of transition between the two pallets placed on each other in the upper right corner, which low-contrast, but perceptible by human observer. Such low-contrast edges/lines are more difficult to recognize for artificial intelligence/automatic algorithms. Perhaps the biggest difference is shown in the rendering of the machine reflected in the puddle: its detail gradually deteriorates when going down from Quality through Balanced, Performance to Ultra Performance mode.

In general, however, the image still retains about equal sharpness. This is a good result. Rendering is less accurate and it is being “manufactured” by artificial computation using a neural network and temporal averaging, but the similar degree of sharpness, equally thin lines successfully disguise that. What probably also helps to mask it is the generous amount of added noise, which I guess might be inserted only after the application of DLSS 2.0 (not sure about that).

Scene 2

In the second scene, the difference between DXR disabled and DXR enabled is shown first (first two screenshots). Notice how the grids/mesh and perforated metal sheet on the wall to the left look quite artificial without DXR. The second image gives them a significantly different look, the metalic and shiny character of the surface look more realistic. Without DXR, this part was unnaturally light (which is probably an attempt to simulate metal surface), but this disappears when DXR is turned on. With ray tracing the shadow under the box at the bottom right looks more realistic too, as expected. However, even the overall contrast of the bright area under the lamp and the darkened parts away from it looks better.

But this scene is also more interesting to us in that you can actually find larger differences between the individual speed settings of DLSS in it. This proves the (arguably intuitive) assumption thatthe  higher FPS can’t be a free lunch and the quality decrease caused by lower native rendering resolution has to show up somewhere.

It can be noticed that when you gradually switch from Quality to Balanced, then Performance and finally Ultra Performance, the sharpness of the text deteriorates somewhat (especially the label for the last weapon at the bottom). Interesting thing to look at is the grid on the left, the three lower panels with the diagonal wire grid – with higher performance (lower quality) DLSS modes, this area suffers exhibts peculiar artifacting: the image of wires going in one of the directions is disappearing and only the the other direction remains, which makes the grid looks like a one-direction stripe pattern. At Ultra Performance, the less illuminated lower fields of this grid are already generally degraded by artifacts. The lines are still being preserved in at least one direction, but they now look somewhat “hairy”. The AI upscaling and temporal filtering is fighting valiantly here, while there is damage, the texture has not blurred/disintegrated completely. But it clearly shows that the visual gradually loses more detail and sustains some damage as you choose these faster DLSS modes. Basically these are the actual costs of that proverbial free FPS lunch.

The detail and rendering fidelity also deteriorates quite a bit on the ceiling between the pipes located there. And what is especially interesting: in this area you can actually see the difference between the best mode DLSS (Quality) and completely native resolution (DLSS Off). You can see the boundaries between the individual narrow pipes partially disappear (which then worsens in faster DLSS modes). There is also quite significant difference in the detail/sharpness of the texture underneath the text, and information shown about the weapon. Even the best DLSS Quality will already make these parts of image visibly less sharp. The difference in rendering quality is quite large here, almost reminiscent of texture blur you can see in video compression. As already mentioned, the higher FPS can not be had for free.

  •  
  •  
  •  
Flattr this!

Leave a Reply

Your email address will not be published. Required fields are marked *