Conclusion

DLSS 3.5 Ray Reconstruction is yet another puzzle piece in NVIDIA's DLSS story. What started with just upscaling has turned into a whole ecosystem of image quality and performance enhancement features for GeForce gamers. At first we had just "DLSS", i.e. Super Resolution, which takes a lower-resolution input and upscales it, better than classic upscaling. A while later, "DLAA" was added, which is DLSS Upscaling without upscaling. Here the image enhancement algorithms work on a native-resolution input, which means no upscaling artifacts, and you still get all the anti-aliasing goodness—we love it! With the GeForce 40 launch, "DLSS 3" Frame Generation was introduced, which looks at two images and creates a third "middle" image, in which things that changed between the two inputs have moved only half the distance. This produces a significantly higher framerate, without taxing the CPU/GPU rendering machinery—it works even when CPU limited.

Today NVIDIA's DLSS 3.5 Ray Reconstruction technology launches, first in Cyberpunk 2077, but many more games are in the pipeline. "DLSS 3.5" is a bit of a weird naming scheme, but it makes sense if you think of "DLSS" as a constant evolution where features get added, big changes bump the version by "1", smaller ones by less, like "0.5" in this case. The fundamental problem is that DLSS means "Deep Learning Super Sampling," which it really isn't anymore. It's actually "NVIDIA's awesome experience enhancers," but I can understand why they wanted to keep the "DLSS" brand, establishing it took a lot of time, effort and money.

From a developer's perspective, things are much more separated. There's now three DLL files to implement DLSS: nvngx_dlss.dll, which contains Super Resolution (and DLAA), nvngx_dlssg.dll, for Frame Generation, and nvngx_dlssd.dll for Ray Reconstruction. These technologies are fairly independent, which also explains the GPU requirements. While Super Resolution works on all GeForce 20 and newer GPUs, Frame Generation requires GeForce 40, yet Ray Reconstruction works on GeForce 20 and newer, just like Super Resolution. This doesn't means that DLSS 3.5 magically enables Frame Generation for older GPUs—a common misconception. According to NVIDIA, Frame Generation requires the optical flow accelerator unit inside the Ada GPUs. AMD has announced FSR 3, a shader-based frame generation technology—it will be interesting to see how well that runs, both in terms of image quality and performance.

Today is all about ray reconstruction though. We've done extensive testing with NVIDIA's new tech and I have to say I'm impressed. For years we've seen ray traced reflections that look grainy and smeared out, but we've accepted that as "that's how it has to be, or performance will suffer (even more)." NVIDIA wasn't happy with that explanation, so they thought long and hard about the problem and came up with a thinking-outside-of-the-box solution. Instead of throwing more rays at the problem, Ray Reconstruction takes advantage of additional information, which is available in the game engine, but lost at some point in the pipeline, because subsequent rendering steps work with less information. Take a look at our comparison screenshots—there really is a significant difference in the final output. I do have to admit that especially in motion and during busy gameplay these differences are not very noticeable, but I still think they improve image quality a little bit, which is why it seems appropriate calling this "DLSS 3.5" instead of "DLSS 4.0." It's not just reflections that look sharper, there's substantial improvement to shadows, too, especially when it comes to fine detail. When the rays are traced, they often "miss" fine-structure detail, because there simply isn't enough rays to sample the structure. With Ray Reconstruction, these objects are resolved much better and shadows look sharper and more detailed. We also noticed improvements to ambient occlusion which looks a bit more realistic. Classic denoisers aggregate multiple frames of information, which means some moving objects like the headlights of a car in motion get smeared out—with RR the headlights are crisp and sharp. The same is happening when the source of a reflection flickers quickly—classic denoisers often average out the effect or show it with a short delay.

Unfortunately Ray Reconstruction is not supported with DLAA or rendering at native resolution. NVIDIA confirmed that they will be training RR for DLAA in the future. What's even more problematic in my opinion is that right now, Ray Reconstruction only works with Path Tracing enabled. This limits the technology to very few gamers—basically those with RTX 4080 or RTX 4090. Everybody else will be playing with classic ray tracing, or even RT disabled. From a technology perspective, nothing stops Ray Reconstruction from working with classic Ray Tracing effects. At the moment, Ray Reconstruction will be limited to a small audience, but once NVIDIA can solve these two problems it will be a game changer for RT in games. I also think that AMD will be able to implement a similar technology for FSR, and it might even help them gain some extra performance, because fewer rays could be used.

During testing I noticed that the image with Ray Reconstruction does look a little bit brighter and some surfaces look a bit washed out—NVIDIA should be able to address these quickly. My biggest troubles with Ray Reconstruction testing were in Cyberpunk's config menus. For some reason the game really likes to mess with your settings. Turning on FG? Here have some "DLSS Auto" with it, your choice of "DLSS Quality" doesn't matter. Oh you're activating DLSS? Then we really want you to see RR in action, here you go, no idea why you chose to disable it previously. Again, nothing CDPR can't fix, but we've seen many such cases before, in which game developers misconfigure their menus for important graphics techniques, possibly due to a lack of time or understanding. At least CDPR got the sharpening sliders right. Having all these technologies at one's disposal is certainly difficult, that's why I feel Microsoft should make a push to unify these technologies under the DirectX umbrella, just like they did with ray tracing. This would make things crystal clear for all parties involved and possibly drive up adoption rates.

An interesting discovery is that Ray Reconstruction actually lowers VRAM usage. For example at 4K we measured 11.8 GB with RR disabled, and only 10.9 GB with RR enabled. One possible theory is that the default denoiser (which gets replaced by RR) has a higher memory usage, possibly because it keeps more history frames in its buffer. What might also help is that RR is integrated with the DLSS Super Resolution pass, which means some buffers can be shared and don't get duplicated.

In terms of performance, NVIDIA was careful with their claims. The official messaging is "it's an image enhancement technique, performance can be similar, slightly better or slightly worse." In our testing we saw single-digit performance gains across the board, especially for Ada and Ampere. For example, the RTX 4090 at 4K without Frame Generation goes from 49 FPS to 53 FPS. I tested a few locations and couldn't find a single one where performance was reduced by enabling RR. This is good news—NVIDIA's new feature not only looks better, it also improves performance. Good job!