After years of hypothesising if or how Nvidia's machine learning-driven DLSS system would work on Nintendo Switch 2, we have finally pieced the picture together. The information is important in analysing present and future Switch 2 software, between ambitious DLSS-enabled ports like Star Wars Outlaws and first-party non-DLSS games like Kirby Air Riders. How does it work? And why might a game embrace or skip any Switch 2-optimised DLSS option?

We begin with estimations calculated in late 2023 using a Switch 2-like Nvidia GPU: an RTX 2050 with 4GB of VRAM running at 750MHz. In these PC tests, we confirmed across the board that enabling DLSS at 1440p or higher resolutions ate up precious per-frame GPU resources, exceeding the 16.66ms needed to reach a steady 60fps refresh rate. Even barring Switch 2-specific optimisations, we determined that DLSS on our testing PC could draw up to a maximum resolution of 1080p, no higher, and still potentially reach 60fps. It was a simulation, of course, but gave some idea of the constraints of a low-power Nvidia Ampere GPU.

Yet within the Switch 2's launch window, we confirmed that some of its games can upscale via DLSS to 1440p or even 4K resolutions and still reach a 60fps performance level, albeit with sometimes harsh image quality deficiencies, not in line with PC DLSS's characteristics. Others at lower frame-rates, meanwhile, enjoy slightly superior DLSS treatment of important elements like sharpness, motion, camera cuts and handling of disoccluded pixels. This latter option does seem to present in a way that has much in common with DLSS as we know it.

After testing games with DLSS implementations on both PC and Switch 2 and analysing our findings, Digital Foundry confirmed with a game developer familiar with Switch 2 development that we were on to something. Switch 2 games can employ one of two DLSS "types," each with their own finer-tuned presets, and we have found examples of each.

The first, which matches up relatively closely with our late-2023 findings, resembles the Convolutional Neural Networks (CNN) model that is used quite widely on Nvidia PC GPUs. It comes with a significant rendering cost and delivers decent Switch 2 results with moving images, as seen in particular on CD Projekt Red's Cyberpunk 2077.

As we understand it, the second only requires roughly half the frame-time cost to render out and is described as a "newer" version of DLSS, though we don't have further information on whether it leans on another CNN-centric pixel-analysis model. It has been described almost identically to what we've found: it is not as capable at smoothly parsing objects and pixels in motion, yet it can provide increased sharpness and clarity in slower motion scenes.

These findings truly pay forward our pre-Switch 2 estimations. The existing state of DLSS, as generally deployed on PC, is too demanding on lower-spec RTX GPUs to efficiently upscale to higher resolutions. That meant Nvidia, Nintendo or a combination of the two had to invent a new DLSS methodology that combined its Tensor core-centric pixel analysis and upscaling with severe hardware limitations.

Based on our tests of games that have deployed this "tiny DLSS" type, we think developers should better appreciate both its strengths in portable mode, where image degradation is less noticeable on a 1080p screen, and its weaknesses in docked mode, where its failure to cleanly account for motion is difficult to forgive at resolutions reaching up to 4K. In short: if you're going to deploy tiny DLSS, limit its use to portable mode, unless your game is optimised around highly static scenery.

That being said, we consider a question from a Digital Foundry user about whether tiny DLSS, with its inherent limitations, is superior as a Switch 2 upscaling option compared to AMD's rival technology, FSR 3, which is a filter-based TAA solution.

From our standpoint, Nvidia's admittedly limited Switch 2-specific DLSS type can be the better option depending on the game. For one, it can lean directly on task-specific Tensor cores in the Switch 2 GPU, instead of employing FSR 3's less optimal use of general GPU cores on Nvidia silicon. That means a game likely spends fewer overall resources using tiny DLSS compared to FSR 3 - so on a cost basis alone, it's more likely to fit into a developer's crucial frame-time budget.

Additionally, we can use FSR 3 as a comparison point to laud tiny DLSS for one of its rendering strengths: superior handling of particle ghosting. FSR 3 is unable to clearly manage particle-based effects, thus leaving lengthy trails behind pieces of dust or drops of rain that we actually haven't seen as much in tiny DLSS. And we keep our fingers crossed that Nvidia and Nintendo come up with further, future optimisations to better translate motion in its tiny DLSS model, which FSR 3 likely will never enjoy to the same extent on Switch 2.

We're not sure we'd honestly recommend either of these two choices for a game like Fast Fusion, whose frenetic racing is far too broken up by motion-based deficiencies to look appealing when combining a full 60fps refresh and high target resolutions. And we see Nintendo's own Kirby Air Riders this week eschewing any machine learning-driven upscaler in favor of SMAA to produce generally clean pixels - thus showcasing its dramatic handling of physically based rendering (PBR) and other impressive effects.

All told though, the team favour Nintendo DLSS over a potential FSR 3 alternative and we'd be inclined to believe that developers agree too, based on the take-up of DLSS up against FSR 3 where both are options for game makers. The question is: when will Nintendo itself tap into the technology with its own first-party studio output?