AMD Universal Upscaling With Radeon Super Resolution, FSR 2.0 Incoming

[harlan4096]

Go native or go home

AMD makes some of the best graphics cards, but even the fastest options can struggle with native 4K rendering. AMD is going all-in on resolution upscaling, with two new approaches. The first is Radeon Super Resolution (RSR), which launched today with the latest 22.3.1 drivers and is available for all Radeon RX 5000-series and new graphics cards — sorry, RX Vega and 500-series owners. We've had early access to the drivers for a few days and have done some preliminary testing, which we'll get into in a moment.

The other big announcement is FSR 2.0, FidelityFX Super Resolution 2.0, which is slated to launch in Q2 2022. Also, AMD's Radeon Software Adrenalin Edition has been rebranded as simply AMD Software Adrenalin Edition, reflecting the fact that it powers both dedicated and integrated graphics solutions, some of which might not carry Radeon branding.AMD FidelityFX 2.0 PreviewWhile FSR 2.0 might seem like the most interesting announcement, we don't have full details on what AMD is doing at present. AMD says that FSR is the "fastest adopted software gaming tech in AMD history," and of course it's always looking to improve on things. Here's what we know about FSR 2.0.

FSR 2.0 will bring improved image quality compared to FSR 1.0 at all upscaling quality presets. This will come in part from a switch from spatial upscaling to temporal upscaling with "optimized" anti-aliasing features. Spatial upscaling means the only data used for upscaling a frame comes from the frame itself. Temporal upscaling in contrast can use data from the current frame as well as previous frames. Interestingly, Nvidia's DLSS 1.0 was a spatial upscaling solution, and DLSS 2.0 switched to temporal upscaling.

That might sound as though AMD will be taking a similar approach to Nvidia now, but there are differences. One key element of DLSS is that it requires large amounts of compute, delivered via the RTX series' tensor cores. AMD doesn't have tensor cores on its current RX 6000 or previous RX 5000 series GPUs, and as the minority provider of graphics hardware, it wants to make FSR work with the widest range of GPUs possible. AMD says FSR 2.0 doesn't require dedicated machine learning hardware and that it will continue to work on products from AMD as well as its competitors.

The above gallery provides the only image quality comparisons so far between FSR 1.0 and FSR 2.0, and if you look at the full size images it's pretty obvious that FSR 2.0 looks better than FSR 1.0 when both are running in performance mode. We'll have to wait for the full release for additional comparisons.

While running FSR 2.0 won't require specialized hardware, it's still possible for it to use AI and machine learning techniques for training the algorithm. That's what Intel's XeSS does, for example, and it can run on Intel Arc GPUs' tensor cores or it can use DP4a (INT8, basically) instructions, which are available in Nvidia Pascal and later, AMD RDNA and later, or Intel Gen11 (Ice Lake) and later GPUs. AMD also hinted that its goals with FSR are to provide substantially more performance than native rendering, which indicates AMD is focused more on a fast algorithm than image quality.

Beyond that, we don't know any other details regarding FSR 2.0, but AMD will host a "Next-Generation Image Upscaling for Games" session at GDC on March 23, where it will provide additional information.AMD Radeon Super ResolutionThe other big announcement is the official launch of Radeon Super Resolution, which was previously revealed as an upcoming technology. RSR uses the FSR 1.0 algorithm — for now, as AMD could potentially upgrade it to FSR 2.0 in the future — and applies upscaling to the entire frame. It's a driver level feature that basically alters the output from your graphics card to your display.

Without RSR, a graphics card can output either a lower resolution signal and leave it to your monitor's hardware to upscale that to fill your panel, or it can internally upscale everything to your monitor's native resolution and send that to your display. With RSR enabled, AMD's GPUs take the latter approach, with the added benefit of using the FSR algorithm to upscale lower resolutions to native.
AMD RSR TestedWe tested out RSR, but immediately uncovered something unexpected: It only supports your display's native resolution. If you have a 4K monitor, which is what I use for all my PCs these days, RSR requires that you upscale to 4K. I wanted to also see how things looked when upscaling to 1080p or 1440p but was unable to do so. Anyway, I tested the Radeon RX 6700 XT using RSR on our usual test suite, since it normally struggled with 4K gaming.

We've got two tables, the first showing the performance loss from enabling RSR (i.e. rendering at the specified setting and upscaling via RSR to 4K), and the second shows the performance gains vs. native 4K at the various settings.

Starting with RSR compared to our normal settings, depending on the game we measured performance losses anywhere from as little as 1% to as much as an 18% drop. It's not clear what causes the difference, as the RSR algorithm ought to be pretty consistent in its requirements. However, games have different requirements and so there may be more or less GPU compute available.

It's interesting that Red Dead Redemption 2, which is the only Vulkan API game we tested, shows a much larger drop in performance. Total War: Warhammer 3 meanwhile shows a consistently smaller hit from RSR, and it's the only DX11 game we tested — the other games all use the DirectX 12 API.

Overall, it's not a significant hit to performance. If you have a 4K display and normally run at 1080p or 1440p because your graphics card isn't capable of handling native 4K at reasonable framerates, RSR can look a bit sharper and more pleasing than letting your monitor handle the upscaling. Now let's look at the potential gains versus native 4K.
...

Continue Reading

[harlan4096]

Additional Info: https://www.techpowerup.com/review/amd-f...fx-fsr-20/