Over the past few years, Nvidia’s Deep Learning Super-Sampling (DLSS) standard has largely delivered on its magical promise: smoother gaming performance and crisper imagery, all based off of zillions of machine-farm computations to predict 3D game visuals. (You can see comprehensive DLSS breakdowns in my reviews of the RTX 3060 and RTX 3080 Ti.) The catch remains that your computer needs a compatible Nvidia “RTX” GPU to tap into the proprietary standard, which has become an ever-tougher pill to swallow in a chip-shortage world.
Still, if you run a DLSS-compatible game on an Nvidia RTX GPU, the performance gains can range from a solid 25 percent to an astonishing 90 percent—usually with greater returns coming from higher resolutions. Up until this week, one demanding PC-gaming use case has somehow not been a part of the DLSS ecosystem: virtual reality.
The default pixel resolution on popular headsets like Oculus Quest 2, Valve Index, and HP Reverb G2 often surpasses an average 4K display, and those headsets also demand higher frame rates for the sake of comfort. Thus, the DLSS promise seems particularly intriguing there. When DLSS works as advertised, a given game renders fewer pixels. This is when Nvidia’s RTX GPUs leverage their “tensor” processing cores to fill in the missing details in ways that, theoretically, look better than standard temporal anti-aliasing (TAA) methods.
Early versions of DLSS needed time and tuning to evolve, but the standard has now become rock-solid on the flat-screen front. Can we say the same about its debut on Wednesday within the demanding VR mode of No Man’s Sky?
“Option 1, or option 2?” “Hmm, let me look again”
The best news comes in the form of the above gallery, which shows No Man’s Sky‘s VR mode rendering while using one of two options (as found in its “anti-aliasing” menu). Which option is which? I’ve hidden the answer so that you can peek at the comparisons and make your own guesses. So keep reading.
One of the options is Nvidia’s DLSS, which offers four levels of detail: quality, balanced, performance, and ultra performance. (Nvidia has yet to clarify exactly what resolution scaling applies to each of those four options.) Base resolution is at its highest at “quality” and lowest at “ultra performance,” and in my cursory testing, I found that “balanced” looked clearer than the lower targets while looking identical to “quality.” So I’ve opted for “balanced” with the above gallery.
The other option in that gallery is base TAA, which works by using frame-by-frame temporal data to smooth jagged edges in a given scene. DLSS requires that a game have a built-in TAA option to apply its own image-reconstruction model to—even though the methods ultimately work differently.
I actually had to rerun my tests at one point because I lost track of which images lined up with which NMS toggle I’d picked. They’re both incredibly similar, and in the above gallery’s case, “balanced” DLSS is option 2, while TAA is option 1. You can find examples of clean lines and smudgy results in both options, though ultimately, I think DLSS wins out as far as NMS‘s VR mode is concerned. And it holds up in action, as well. NMS‘s TAA adds a certain ungainly flicker to various swaying objects and moving clouds, which DLSS happens to resolve neatly, all while adding additional performance. (I’ll get to that later).
Now, you may expect to get better performance in another non-Nvidia way: by disabling TAA and switching to no anti-aliasing (AA) at all. (DLSS, as a reminder, is a mix of antialiasing and image resampling.) As it turns out, that’s not quite the case; NMS’s TAA runs at only a tiny cost higher than no AA at all. Above, I’ve included a direct comparison of the same scene with all AA disabled against the scene with DLSS enabled at “balanced” mode. The visual difference is immediately apparent, thanks to the jaggies in the first image. DLSS, again, runs faster here, while offering smoothing wholly comparable to what you get from turning TAA on.
Ranging from 11-20 percent higher
As far as my testing rig’s performance, meanwhile, DLSS’s VR gains are noticeable but not life-changing. With an RTX 3070 strapped into my rig, I found that, for every step I activated in DLSS’s fidelity scale, I managed to claw back 6.25-8.5 fps in performance. Booting the game in default settings with TAA activated got me, as of press time, close to 105 fps in a randomly generated biome on the game’s latest patch. Then I toggled DLSS’s “quality” mode, which saw the frame rate jump to 111.25 fps, then kept bumping down on the fidelity scale, seeing 117.5 fps in “balanced” DLSS and 123 fps in “performance” DLSS.
Meanwhile, I went back to do the above “all AA disabled” tests in a different biome, and the results were more staggering in “balanced” DLSS’s favor, to the tune of 20 percent gains over either TAA enabled or all AA disabled. But I could not regularly reproduce that larger performance gap between DLSS and other options.
As I previously mentioned, “balanced” got me where I liked on a visual basis, which amounts to a roughly 11 percent higher performance level over default TAA, all without an apparent hit to quality or stability. Worth noting, DLSS does not magically fix the game’s greater issues of scaling from a flat-screen adventure to a comfortable VR game. My testing rig (i7-8700K, 32 GB DDR4 RAM, SSD) still suffers from frequent frame-time spikes, whether due to procedural planet generation or asset-loading hitches. But DLSS does help with across-the-board frame-rate increases.
Hopes for DLSS-VR’s future?
To clarify, if you like NMS but don’t care about its VR mode, you can also cash in on DLSS’s impact in traditional, flat-screen gameplay. This week’s update adds DLSS to every PC mode imaginable for Nvidia card owners.
I had a couple of big questions going into my latest tests. First, how does DLSS work when it has to touch up two simultaneous images, one for each eye? And second, would its existing workflow adapt to that double-image frenzy? Maybe that’s why this system’s gains seem meager compared to some of DLSS’s gains in other benchmarks, but at its lowest boon, a raw 11 percent gain is nothing to sneeze at, especially when VR gamers look for every way imaginable to lock higher-end games closer to a 90 fps refresh. (If your own testing scenario gets closer to 20 percent gains, like mine revealed from time to time, then that’s even better.)
Nvidia has yet to reply to my questions about DLSS’s current VR efficiency. Meanwhile, a Hello Games rep responded to Ars’ questions by pointing their finger at Nvidia. Still, even while major details continue to hide behind Nvidia’s veil, DLSS has finally proven that it does improve a VR game’s performance—and we hope Nvidia spreads this gospel with as many major VR developers as possible in the near term, especially as the PC-VR space continues to slip compared to lines like the Facebook-dominated Oculus Quest ecosystem.