Deathloop: FSR 2.0 vs FSR 1.0 vs DLSS 2.3 in Benchmark and Image Comparison

The time has come: with Deathloop we have the first game to support AMD’s FidelityFX Super Resolution – or the second iteration of FSR 2.0. FSR 2.0 should now also include temporary data, but does not require an AI network. Since AMD doesn’t use any AI accelerators in its current GPUs, this shouldn’t come as a surprise. AMD also sees disadvantages in using AI in this area.

Apart from benchmarks, today we will also look at the image quality of the mode in Deathloop. In the future, users of AMD graphics cards will have the choice between a native display and (possibly) two FSR modes, with FSR 2.0 likely to replace the first generation in the long term. In the future, owners of an NVIDIA graphics card can choose whether they want to use FSR 2.0 or DLSS (possibly even FSR 1.0), as long as the game supports both upscaling technologies. Same is the case with Deathloop.

Like FSR 1.0 or DLSS, FSR 2.0 also offers different modes. These are described as “quality”, “balanced” and “performance” and also reflect the focus set out here.

Depending on the desired output resolution and the selected quality level, a correspondingly lower rendering resolution is used. The scaling factor is between 1.5x and 2.0x. This reduction in resolution leads to an increase in the performance of FSR 2.0 – and FSR 1.0 and DLSS. Sharpening filters and AI reconstruction (in the case of DLSS) restore image detail and sharpness.

Benchmark:

First, let’s look at the performance of different GPUs with FSR 2.0. On the ASUS ROG Maximus Z690 Hero an Intel Core i9-12900K with DDR5-5200 is used as a test system.

But how the performance of a GeForce RTX 3080 and Radeon RX 6950 XT with FSR 2.0 evolves in different modes. We have also investigated this question.

A GeForce RTX 3080 increases 33% from “native” to “quality”, 63% to “balanced” and 90% to “performance”—nearly double that. The Radeon RX 6950 XT comes in from “original” to “quality” for a plus of 38 and 51%. We speak of 64% for the “Performance” profile.

Next, the comparison between FSR 1.0 and FSR 2.0, which is certainly not uninteresting.

Basically, an FSR 1.0 is faster than the newer FSR 2.0 version. In the “Quality” preset we’re talking about 12%, for “Balanced” we’re still talking about 13%, and for “Performance” we’re talking about 16%. So using FSR 2.0 will yield a lower FPS than using FSR 1.0. Of course, we’ll also see if the higher image quality can compensate for this.

Finally, along with the benchmarks, we want to take a look at the direct comparison between FSR 2.0 and DLSS 2.3 in different modes:

An FSR 1.0 is always sharp here, but based on past experience it also delivers the worst image quality. The new FSR 2.0 and DLSS 2.3 are on par in terms of performance.

image comparison

Performance is a factor for an advanced technology. Image quality comes in second – and in the past there were often big differences between FSR 1.0 and DLSS – mostly in favor of NVIDIA’s technology, which puts in significantly more effort to achieve image output.

With FSR 2.0, AMD now also has a temporal component – in other words, the information from the previous frame is also taken and taken into account. Thus, the image should be much more stable and the flicker of various objects is reduced.

First we look at the display quality between DLSS 2.3 and FSR 2.0, along with native rendering in the “Quality” preset. We did native rendering without any anti-aliasing, so the rough steps can definitely be seen here. The DLSS image looks a bit sharp but also shows slight jiggling, while the FSR 2.0 can smooth them out fairly well without completely blurring.

A second comparison like this, on the other hand, again shows a greater benefit for performance using DLSS, as the shadows around the frame on the shop window show significantly fewer “pieces” of anti-aliasing missing in the native rendering, Which is also reinforced by the low rendering resolution.

A comparison of the “Quality” presets for FSR 1.0, FSR 2.0, and DLSS most clearly shows the evolution of FSR 2.0. While FSR 1.0 apparently produced a worse image with missing details than DLSS, these are largely available again with FSR 2.0. But even here, the assumption prevails that the image with DLSS is sharp but has phases to show for it, and the image with FSR 2.0 is “smoother” but appears less sharp.

In the first image comparison, we compare the “Balanced” presets of DLSS and FSR 2.3. In the parabolic antenna and the bottom fence, many of the details that still exist with DLSS are lost in calculations using FSR 2.0. In terms of sharpness, both the screenshots are on the same level. The level of detail is even worse for the “Performance” preset in comparison to the second image. NVIDIA clearly still has some advantages, thanks to rebuilding options that AMD doesn’t have.

conclusion

Advanced technologies were created out of necessity as well as a useful tool. In addition to the sometimes extremely complex ray tracing effects, this ensures that even high-end cards are not capable of delivering enough FPS. Thus, a display in UHD with all the details should still be possible. Of course, a 1080p card isn’t going to match the performance of a UHD display model, but if it doesn’t have enough performance to get smooth FPS at the desired resolution, upscaling techniques may make sense.

DLSS has seen strong growth ahead, which has been continued and has taken a huge leap forward with version 2.0. FSR, on the other hand, has remained largely untouched since its introduction and is only now taking a big leap again with version 2.0.

It is not so easy to compare the techniques, as they all work in different ways and hence each has its own advantages and disadvantages. FSR 1.0 implements a lower rendering resolution and uses a sharpening filter for an enhanced image. DLSS added a temporary component from version 2.0 and also uses missing detail reconstruction using AI. Depending on how well the neural network works, the details that were lost due to the low rendering resolution may also be added here. FSR 2.0 now also adds a temporary component, but omits AI reconstruction.

What matters in the end is what comes out. FSR 1.0 could not sync with DLSS. There were cases where it worked quite well, but often quite poor image quality could be quickly identified. But DLSS was not good from the beginning, it also had to be further developed earlier, which took months. However, FSR 2.0 has now reached a level that is not yet quite close to DLSS (2.3), but also comes without the limitations of hardware. For game developers, it should be an important argument that FSR 2.0 works on hardware with both AMD and NVIDIA GPUs – and without any difference in performance quality. DLSS, on the other hand, only runs on GeForce RTX cards. This doesn’t seem technically necessary, and NVIDIA may offer a fallback solution by computing via shaders, but hasn’t done so yet. Intel’s XeSS also won’t use AI reconstruction and will benefit from AI accelerators, but here’s at least the intention that XeSS will work even on hardware without such specialized hardware.

FSR 2.0 is an important step forward for AMD. The increase in performance is slightly less than that of FSR 1.0, but the quality of the display increases significantly. In the end, the user himself has to decide what he likes better. It is always a compromise between display quality and performance. The static images and benchmarks we show can only be a guide and depend on a number of different points of view.