Conclusion
There you have it, our look at FSR 3 analytical Upscaling and Frame Generation performance vs. AMD FSR Redstone ML-Upscaling and ML-Frame Generation. We tested the AMD Radeon RX 9070 XT, Radeon RX 9070, and Radeon RX 9060 XT across 1080p, 1440p, and 4K in the games listed on the AMD gaming list that show support for FSR 4 ML-Upscaling and AMD FSR Redstone ML-Frame Generation.
FSR Upscaling Performance
To sum up performance, it’s really as easy as saying that ML-Upscaling will cause a performance impact versus FSR 3/3.1 Upscaling to your AVG and 1% Low framerates. The impact, however, on the Upscaler component is minimal in games, but it does exist. What it will affect more is the 1% Lows in games that are very demanding, and already performing slowly even with FSR Upscaling enabled.
There could be cases, as we experienced, where the new AMD FSR Redstone/FSR 4 ML-Upscaling drops the 1% Lows below 60FPS, whereas FSR 3 Upscaling kept it at 60FPS or above, just due to the small 5% on average performance impact. There was one outlier in Dying Light: The Beast, where it had a 12% impact to 1% Lows, but generally it was around the 4-5% range. It will most likely be game depenendent, and even resolution-dependent depending on what native resolution you are starting from.
FSR Frame Generation Performance
When it comes to AMD FSR Redstone ML-Frame Generation, which is the new component launched this December, we find that it also impacts performance compared to FSR 3 Frame Generation. The new ML-Frame Gen is consistently slower than FSR 3 Frame Gen, and it varies more than ML-Upscaling does, with a deeper impact overall since it drops the 1% Lows further down. We saw ML-Frame Gen generally be closer to 7% on average, slower than FSR 3 Frame Gen.
But we did see greater differences in some games, with some cards, like the Radeon RX 9060 XT in CoD at 18% slower 1% Lows, and in Cyberpunk 2077 and Hogwarts, it has a 9% impact. There might be an indication in all of this that the Radeon RX 9060 XT is impacted more by ML-Frame Generation performance impacts in games.
One thing is true when it comes to ML-Frame Generation, it shares the same fate as any Frame Generation technology in that input is not run through the game engine pipeline, it doesn’t account for it, but insteads add latency without input. Therefore, it works as a smoothing technology. As we can experience from the benchmarks in this review, in many games the 1% Lows remain close to the 1% Lows of Upscaling, with no improvement, or any improvement that matches the increase in AVG framerate. It creates a scenario where the 1% Lows and AVG FPS have a very large gap, with a lot of FPS deviation between them. This is opposite Upscaling, which keeps the distance between 1% Lows and AVG close, since it includes game engine input.
To put it another way, let’s say you are getting 60FPS 1% Lows, and 70FPS AVG, that is a difference of 10FPS. Now, say you turn on Frame Gen, and the 1% Lows remain at 60FPS, but the AVG has increased to 100FPS, the difference is now 40FPS. The gap is wider. That creates a sluggish feeling sometimes, lag, lower latency, and the general “feel” is off. This is a very oversimplified explanation, but we hope you can see it in the benchmark graphs. Upscaling doesn’t do this because the 1% Lows scale with the AVG FPS gain.
FSR Ray Regeneration Performance
Overall, AMD FSR ML-Ray Regeneration decreases performance slightly compared to the Default Denoiser in Call of Duty: Black Ops 7. This is opposite NVIDIA DLSS Ray Reconstruction, which, so far in our testing, actually benchmarks slightly faster than the game’s default denoisers. It is important to note that the goal and point of Ray Reconstruction and Ray Regeneration is not to improve performance; the goal is to improve image quality. It is just a happy net positive that DLSS Ray Reconstruction actually improves it slightly, and looks better.
However, for AMD, it seems FSR Ray Reconstruction consistently on each video card, at different resolutions decreases performance by 7-8% on AVG and 6-7% on 1% Lows, which is significant. It is especially significant knowing how much of a performance hit Ray Tracing already takes in games, which renders Call of Duty: Black Ops 7 unplayable without FSR Upscaling.
All of that said, this is but one game with ML-Ray Regeneration support; we will have to wait and see how other games behave when/if they get support.
Final Points
AMD FSR Redstone technologies, including ML-Upscaling, ML-Frame Generation, and ML-Ray Regeneration, are great improvements from the previous version of Upscaling and Frame Gen. Though ML-Upscaling does have a performance impact versus FSR 3 Upscaling, the benefits far outweigh that minimal performance impact. If you have the option to run ML-Upscaling vs FSR 3 Upscaling, you should absolutely do so; the image quality benefits justify the performance difference.
The same is also said for ML-Frame Generation; if you are going to use Frame Generation, you should opt to use ML-Frame Generation as the benefits justify the performance difference. Finally, AMD has ML/AI-(Whatever)-Powered Upscaling and Frame Generation. It has been a long time coming; one can argue AMD should have had this debut with the launch of RDNA 3, but at least, finally, we are getting it with RDNA 4.
The only issue with AMD FSR Redstone is the staggered, disjointed launch of components and the integration and adoption rate in games. ML-Upscaling is finally being patched into games, and we are seeing growing support of games actually having in-game graphics menu options for “FSR 4.” In the games that support FSR 3.1, we can force ML-Upscaling from the AMD Adrenalin Software control panel. For ML-Frame Gen, it’s a lot less supported, with no in-game options at the moment. The only way to force it is through the driver control panel, but only if the game itself supports the latest FSR 3.1.4 version of Frame Generation, and not as many do support that latest version.
Then there is AMD FSR ML-Ray Regeneration, while technically ‘launched’, it is only in one game, one, and it isn’t even usable in that scenario of the game, because who would play with RT in multiplayer? This is the main problem with these technologies: adoption and integration. We must wait on game developers to either patch in “FSR 4” upscaling into the game with menu options, or update the games to support AMD FSR Redstone ML-Frame Gen in-game, or at least FSR 3.1.4 Frame Gen. With Ray Regeneration, we must wait for game developers to integrate that feature.
For the last feature of AMD FSR Redstone, it isn’t even in any games out right now, AMD Radiance Caching, so that is yet to be seen. So, can you really say AMD FSR Redstone is fully released right now? This staggered, disjointed nature of AMD FSR Redstone has made it very frustrating for gamers. Gamers don’t like technologies announced, but with absolutely no way to use them. So we will wait and see.
As a quick aside, just from a subjective viewing experience using ML-Upscaling and ML-Frame Gen, I personally thought it looked great visually. There were a few games where ML-Upscaling has maybe worse aliasing/shimmering issues compared to FSR 3 Upscaling. In Hogwarts, we noticed pretty bad shimmering on grass and trees as we walked through or viewed trees at a far distance. There were some other games where it seems ML-Upscaling has a problem with these kinds of objects in games, where the blurred nature of FSR 3 actually looks better. But for everything else, ML-Upscaling is great, and a big improvement over FSR 3.