Friday, 27 May 2022

Deathloop: Putting AMD’s FidelityFX Super Resolution 2.0 to The Test Against Nvidia’s DLSS

Earlier this month, AMD offered us the chance to preview FidelityFX Super Resolution (FSR) 2.0, courtesy of the game Deathloop. Deathloop is currently the only title to support both versions of FSR as well as Nvidia’s Deep Learning Super Sampling (DLSS), making this an excellent opportunity to take them both out for a collective spin.

Despite the name, FSR 2.0 is not an update to FSR 1.0. It’s an entirely new approach that’s designed to bypass some of FSR 1.0’s weaknesses. AMD’s first attempt at this kind of upscaling was based entirely on spatial upscaling and did not use temporal information. It also required a high-quality anti-aliasing method to work properly, which many games don’t support. Deathloop lacks this support, and FSR 1.0 in the game isn’t much to write home about.

AMD is more confident in FidelityFX Super Resolution 2.0 than it ever was in FSR 1.0. FSR 1.0 was positioned as an alternative feature along the lines of Radeon Image Sharpening, while FSR 2.0 is positioned as more of a direct competitor to DLSS.

FSR 2.0 incorporates the temporal data that FSR 1.0 lacks and it doesn’t require a game to support high-quality antialiasing in order to render acceptable output. AMD previewed it back in March, but this is the first time we’ve gotten to play with it. We’ve taken FSR 2.0 out for a spin against DLSS on a 1440p panel to capture the rendering differences.

One thing to note before we get started. The small gray spots you see on some images are not errors introduced by any of these visual settings. They’re transient phenomena. There’s one place where FSR and DLSS both introduce errors in our comparison and we’ll call it out when we get to it.

How to Compare DLSS and FSR in This Article

This article contains a mixture of directly embedded images as well as links out to Imgsli. Imgsli is an excellent free method for comparing two or more images in an A/B(C) method. Because of the number of comparison points, we’re going to queue comparisons in DLSS, FSR 2.0, and then directly head-to-head for our selected scenes. You can select which image you compare in Imgsli using the drop-down menu at the top of each image.

Nvidia DLSS: Beach

Let’s get started. Deathloop begins with your own murder, after which you wake up on a deserted beach. Here’s the zoomed-out, native 1440p version of the image with no AA or alternative upsampling:

Nvidia Native versus DLSS. Adjustable image comparison available at Imgsli.com

But zoomed out doesn’t give us the best view of what changes in each scene. It’s actually hard to tell what’s different across these frames. (From Nvidia’s perspective, that’s a good thing). A 600 percent zoom is a much better way to see rendering subtleties.

Nvidia DLSS quality comparison. Close-up of shot above. User-adjustable image available on Imgsli.

Quality DLSS settings substantially reduce jaggies compared to the 2560×1440 native resolution. This is expected — DLSS performs antialiasing, in addition to its other functions — but the difference is large. Shifting down to “Balanced” hardly impacts image quality at all. One downside to DLSS (and this is present in every mode) is that ground textures are a bit blurred compared to the native image. This is really only visible at tight zoom, however.

AMD FSR 2.0: Beach

According to AMD, FSR 2.0 is better than FSR 1.0 at every quality level. We focused our testing on FSR 2.0 for these AMD comparisons, but include some FSR 1.0 shots as well, to show the degree of uplift.

AMD native 1440p versus FSR 2.0 versus FSR 1.0. User-comparable results available on Imgsli.

The improvement from FSR 1.0 to FSR 2.0 is immediately obvious. FSR 1.0 blurs content heavily and the line leading away from the pole is a vague smear. With FSR 2.0, it resolves into a distinct line. Switch to “Performance” for both tests, and you’ll immediately see how much better FSR 2.0 is compared to FSR 1.0. AMD claimed that every FSR 2.0 quality setting was better than every FSR 1.0 quality setting, but this comparison shows AMD is actually underselling its own feature. Even FSR 2.0’s “Performance” setting is better than FSR 1.0’s “Quality,” though Deathloop isn’t considered a great test case for Fidelity FX Super Resolution’s first iteration.

AMD 1440p native versus FSR 2.0 and FSR 1.0. User-adjustable comparison available at Imgsli.

FSR versus DLSS: Beach

When it comes to FSR 2.0 versus DLSS, FSR 2.0 wins the comparison in this set of images. Note: We’ve combined the standard shot and closeups in this comparison to try to keep the amount of clickable material to some kind of reasonable limit.

AMD FSR 2.0 versus Nvidia DLSS. User-adjustable comparison available at Imgsli.

FSR 2.0 is much less blurry than DLSS, at every detail setting. We’ve included both the zoomed-out and zoomed-in shots to illustrate the distinction in both modes. FSR 2.0’s “Balanced” preset offers better image quality than DLSS’ “Quality” preset. One thing we do encourage you to keep in mind is that the relative quality of DLSS and FSR can vary considerably depending on the suitability of the game engine for the format and the amount of work invested by the developer. These comparisons might play out differently in another title.

The gains from FSR 1.0 Ultra Quality to FSR 2.0’s “Quality” mode are quite impressive. Even at top quality, FSR 1.0 struggled to distinguish the wire strung up at the pole from background clutter, and lower-quality versions of the feature all but lose the strand. One of AMD’s promises for FidelityFX Super Resolution 2.0 was that the feature’s “Balanced” mode would be better than FSR 1.0’s “Ultra Quality” mode. In some ways, FSR 2.0’s “Performance” mod is better than UQ FSR 1.0, though we wouldn’t actually recommend using Performance mode.

Nvidia DLSS: Bunker

Let’s move from the beach to the interior of the starting area and check out a nearby underground bunker.

Nvidia native rendering versus DLSS. User-adjustable image comparison via Imgsli.

Nvidia DLSS and AMD’s FidelityFX Super Resolution both create a weird textured problem on the ground in this scene. You might not know this was an accident if you didn’t look closely at other rendering modes — while it’s a bit odd looking, the texture doesn’t flicker or change dramatically as one moves around the room. Light across the scene is a fair bit different between 1440p with no DLSS and DLSS engaged, but you can see how DLSS prevents horizontal line shimmer where there are lines across surfaces.

Apart from the introduced error, I consider DLSS Quality to improve the overall image (and FSR also creates the same error). DLSS Balanced, on the other hand, not so much. It’s not that DLSS Balanced doesn’t have any advantages over native 1440p, but there are trade-offs as well, especially considering the damage. We’ll look at a few of these when we zoom in. Temporal AA is the best quality of all, if only because there’s no error on the ground.

Our bunker close-up shot focuses on the map board at the back of the room. It’s striking how bad the default native rendering is. From our vantage point in front of the orange tarp, the close-up native line isn’t actually a solid line of string at all, but a series of dashes. DLSS Quality fixes both the dashes and the detailing on the metal box to the left of the map. DLSS Balanced and DLSS Quality are quite similar here.

Nvidia native rendering versus DLSS, bunker close-up. User-adjustable image comparison available on Imgsli.

Interestingly, Temporal AA is worse on this map closeup than the other settings, even if it looks better in the scene as a whole. Line weights and handwritten text on notes pinned to the board are both stronger with DLSS. Temporal AA manages to beat native, but the setting does not impress here.

AMD FSR: Bunker

The bunker on an AMD GPU has the same visual problem that DLSS has. Both DLSS and FSR change how shiny certain surfaces are, and how reflective they look. It’s not a bad thing, but it does stand out as a difference between enabling and disabling these technologies, even if the floor wasn’t oddly textured. The problem, despite being quite visible, doesn’t really stand out in gameplay with FSR, either.

AMD FidelityFX Super Resolution versus native resolution. User-adjustable comparison available on Imgsli.

Pan back and forth in the image comparison above between native 1440p and DSR 2.0 Quality, and you may notice that one of the lockers in the back appears to lose a line that defines an upper compartment. Zoom in, and it’s easier to see that while the line wasn’t removed, it shimmers less and is less visible. You can also see that FSR 2.0 improves the string rendering on the wall map in the back, even without zooming. FSR 1.0 Ultra Quality looks somewhat worse than native 1440p with no AA technology enabled.

Comparison between native AMD rendering and FSR 2.0. User-adjustable comparison available on Imgsli.

Not much new to say here. Native looks bad on AMD as well, and FSR 2.0 is substantially better than FSR 1.0. I forgot to grab a “Balanced” screenshot for FidelityFX Super Resolution 2.0 for this one — my apologies. But this is an easy win for FSR 2.0, without much more to say about it.

FSR versus DLSS: Bunker

Nvidia DLSS versus AMD FSR 2.0. User-adjustable comparison available on Imgsli.

Both companies’ solutions create an error on the floor, so we’re going to call that a wash and compare on the basis of other characteristics. You may be hard-pressed to see much in the way of variance unless you zoom in, at which point some distinctions appear. Once again, FSR 2.0 is a slightly sharper solution while DLSS blurs just slightly more. Differences this small typically come down to personal preference — do you like a bit of blur to guard against shimmer and jagged lines, or do you prefer maximum detail?

DLSS versus FSR 2.0: Bunker Close-Up

Neither DLSS nor FSR 2.0 look fabulous in this close-up shot, but DLSS gets the nod from us for its ability to create slightly more legible text. Line strength is better with FSR 2.0 compared to Deep Learning Super Sampling, but we’d give the nod to Nvidia overall.

User-adjustable image comparison available via Imgsli.

Nvidia DLSS – Panel Close-Up

We’ve pivoted (literally) towards the console panel you can see above, to get some close-up shots and measure DLSS versus FSR at minimum range. We’ll start with the DLSS comparisons, though we’ve also chucked an Nvidia run of FidelityFX Super Resolution into the mix, just to see how an Nvidia card fares when using AMD’s older rendering method.

Nvidia Native 1440p versus DLSS Quality. User-adjustable image comparison available at Imgsli.

DLSS Quality looks quite similar to native resolution here. While there’s a slight blurring, it’s not very much. AA methods often create at least a small amount of blur, after all. Balanced quality is noticeably worse, however, with significant gauge blur and fine detail loss. Temporal AA deals with some bright jagged lines that DLSS Quality doesn’t and changes the overall lighting a bit. FSR 1.0 does a reasonable job cleaning up the image in some places, but it creates text distortion in the gauge readouts.

Nvidia running FSR 1.0 versus AMD. Image isn’t perfectly aligned due to the need to swap GPUs. User-adjustable image comparison available at Imgsli.

Here, the slight blurring from DLSS Quality is preferable to the increased jaggies in the FSR 1.0 image. FSR 1.0 isn’t really the point of this article, but we wanted at least one comparison between Nvidia and AMD on this point. While FSR 1.0 output isn’t literally identical between the two companies — AMD’s text on the panels is ever-so-slightly blurrier than Nvidia’s — the two are close enough to demonstrate equivalent support.

AMD FSR 2.0: Panel Close-Up

Here’s AMD’s close-up on the instrument panel, compared across native resolution, FSR 2.0, and FSR 1.0.

Native resolution versus FSR 2.0 versus FSR 2.0. User-adjustable comparison available at Imgsli.

FSR 2.0 really shines here. The panel is higher quality with less blurring with FidelityFX Super Resolution 2.0 enabled in Quality mode than it is in native 1440p, as shown below:

A zoomed-in comparison of the two images shown above.

FSR 2.0 improves AMD’s image quality over and above baseline. That’s a trick FidelityFX 1.0 can’t match.

AMD FSR 2.0 versus DLSS: Panel

AMD’s FSR 2.0 wins this comparison against DLSS. The sharper rendering DLSS 2.0 offers pays dividends here, rendering written text and gauge numbers easier to read compared to DLSS. DLSS, in turn, renders significantly better text than FSR 1.0. Both technologies perform well here and the gap between them isn’t huge.

DLSS vs. FSR. User-adjustable comparison available on Imgsli.

While we preferred DLSS for the background map and text in our previous comparisons, we like FSR 2.0 more for the panels and associated gauges.

Putting It All Together: Who Comes Out on Top?

Between DLSS and FSR 2.0, I narrowly prefer FSR 2.0. Honestly, it’s a wash at anything less than a painstaking comparison — it’s not as if you notice a fractional difference in text that’s too blurry to read when playing the game normally. Both technologies broadly deliver what they say they will — namely, a performance improvement even at the highest quality settings.

What matters more for AMD is matching Nvidia’s ability to field an image-enhancing algorithm that improves performance instead of hurting it. In that regard, FSR 2.0 succeeds tremendously.

Technologies like FSR 2.0 could be particularly helpful to mobile and low-power device gaming, especially on products like the Steam Deck. Tests show that technologies like DLSS and FSR can improve rendering performance by 20 – 40 percent depending on the title and your preferred settings. Improving performance this much typically requires buying a new GPU at a substantially higher price.

This shift has short-and-long term implications. Because FSR 2.0 requires RDNA2 support, unlike FSR 1.0, the number of people who can take advantage of this technology is small. Over time, however, this feature will be a mainstream capability in every GPU that AMD manufacturers. Intel will presumably follow suit. Once that happens, gamers can look forward to substantially better performance.

Long term, we expect Intel, Nvidia, and AMD to shift their efforts towards a mixture of AI and non-AI techniques intended to improve image quality without paying the penalty of rendering pixels at their native resolution. FSR 2.0 is an important step on that journey.

Now Read:

 



No comments:

Post a Comment