When the PC industry's two biggest graphics card manufacturers aren't battling over benchmarks or chip shortage woes, they've been fighting over a different sales pitch: boosting performance for older GPUs. And while Nvidia has largely won that war, that has come with an asterisk of a proprietary performance-boosting system, DLSS, that requires relatively recent Nvidia hardware.
AMD's first major retaliatory blow came in the form of 2020's FidelityFX Super Resolution, but this open source, hardware-agnostic option has thus far proven inadequate. And AMD finally seems ready to admit as much in its rollout of FSR 2.0, which debuted in limited fashion on Wednesday ahead of a wider Game Developers Conference reveal next week and a formal rollout in video games starting "Q2 2022."
It’s time for temporal solutions
Both FSR and DLSS function in modern games as pixel upscalers. In both cases, games run at a lower base resolution of pixels, and then, whichever system is active processes and reconstructs the resulting imagery at higher pixel counts. This can include intelligent anti-aliasing (to reduce "stair-stepping" of diagonal lines), blurring, or even wholly redrawn pixels. Ultimately, the dream is that these systems can wisely convert games running at 1440p or even 1080p to something nearly identical to a full 4K signal.
On the most basic level, DLSS's reconstruction system leans upon a dedicated core in Nvidia's "RTX" line of GPUs, which has been trained on a machine-learning model that has studied thousands of hours of existing video games. DLSS uses additional data to reconstruct its imagery, as well, but its dedicated upscaling core has been the difference maker for games that might otherwise be tricky to upscale. Remedy's 2019 hit Control remains a standout example of DLSS often looking superior to native, higher-res pixel counts, especially in how it reconstructs complex textures and in-game text.
AMD can't always compete with that aspect if it leans on a hardware-agnostic upscaler, because it's otherwise at the mercy of whatever base pixels are available from a game itself. As AMD now admits on its GPU Open news site, "with a low-resolution source, there is just not enough information with a spatial upscaler for thin detail."
That's where AMD FSR 2.0 comes in—or, at least, AMD wants us to think so. The update's biggest addition is a juggle of temporal data (motion vectors, movement within a game, etc.) as a double-edged tool: to anticipate and draw pixels and to replace traditional temporal anti-aliasing (TAA). If that system upgrade sounds familiar, it should, as DLSS includes its own temporal pixel interpretation system. Additionally, AMD hints that its system can treat a game's color data as its own vector to process and upscale. Neither Nvidia nor AMD has publicly described how color-mapping upscale systems work, and we're intrigued by this aspect being hinted at loudly by AMD this week.
So far, the only promising example of the new system in action comes courtesy of a YouTube video, which shows the FSR 2.0 system more nimbly translating temporal data as the first-person perspective shifts in the 2021 Bethesda game Deathloop. In good news, from what we can tell, FSR 2.0 appears to reduce the blur and pixel dithering that we saw in FSR 1.0's ugliest in-motion examples from the past year.
Limited evidence thus far—and RSR in the meantime
Sadly, AMD did not make this video's raw assets available as part of today's announcement, which leaves us poking at frozen images as captured from a compressed YouTube feed to cast our initial judgments. Worse, the video never allows viewers to directly compare how each visual option (native pixels, FSR 1.0, FSR 2.0) looks; instead, it splits the screen in half or thirds, forcing us to compare completely different parts of the gameplay sequence in question. There's a point in the above embedded video at roughly the 1:05 mark where FSR 1.0 produces an unsightly flicker next to a light source, and AMD doesn't let us see exactly how FSR 2.0 translates that part of the image.
AMD has also confirmed that FSR 2.0 will require per-game implementation by game developers, and it's unclear thus far whether that's a few hours' labor that piggybacks off a game's existing TAA implementation or something far more intensive. (Though I'd like to hear AMD answer game makers' questions at GDC next week on that front, I fear that those answers may be hand-waved until FSR 2.0 is closer to launch.) And we're left wondering exactly how much in-game performance will improve with FSR 2.0 enabled for supported GPUs, since it will likely require some lifting on the GPU's part, much like how traditional TAA taxes a GPU to some extent. To that end, we don't yet know how much FSR 2.0's updates will narrow device support, at least in terms of spec requirements like VRAM or boost clock speeds.
Still, Deathloop's promised inclusion in the FSR 2.0 launch will be an intriguing one to watch, as it's already a rare example of a game shipping with both FSR and Nvidia DLSS built into its PC version—which will help us prime our own head-to-head image comparisons down the road. And this future-minded news is joined by AMD launching a game-agnostic launch of "Radeon Super Resolution," a driver-level toggle available starting today for AMD GPUs that will upscale lower-res games that otherwise don't have custom FSR support baked in. Because of its lack of game-specific coding, AMD admits that RSR will look inferior to FSR when both options are available, but for anyone inclined to reduce in-game resolutions for better frame rates anyway, RSR is an interesting option to lean on as opposed to simply rendering fewer pixels.