AMD has finally brought a high-quality upscaling solution to RDNA 2 GPUs. That's right, FSR 4.1 is now available for the Radeon RX 6000 series of graphics cards, as well as a range of other GPU architectures, including RDNA 3.5 APUs. Originally, we were expecting to see FSR 4.1 on RDNA 2 at some point in 2027, but with the latest 26.10.1 driver release, it's arrived ahead of schedule, and I'll be fully testing it today. To bring FSR 4.1 to RDNA 2 graphics, AMD claims they went back into the model to tune it and optimize it for the entire RX 6000 range. Both RDNA 2 and RDNA 3 utilize an int8 version of FSR 4.1 upscaling, as opposed to the full FP8 version that runs on RDNA 4. Architectures prior to RDNA 4 don't support FP8 hardware acceleration, so AMD needed to port the FSR 4.1 model to the slightly lower precision int8 format, and it sounds like further optimizations have been made specifically for the RDNA 2 release.
Whether or not those optimizations affect image quality, we'll find out shortly. In addition, AMD are releasing another version of FSR 4.1 specifically for some RDNA 3.5 APUs, typically used in handhelds and some laptops. This version uses what AMD claims is a highly efficient model that's half the size. I'm not going to test that model today. I'm not sure how similar it will be to FSR 4.1 int8, but it is available starting today. The lack of a decent upscaler for RDNA 2 hardware has been something that has frustrated me for ages. FSR 3.1 is just not good enough, especially when used at resolutions like 1440p, and it's broadly uncompetitive with Nvidia's DLSS. This has led to a huge feature disparity between RX 6000 GPUs and RTX 30 series GPUs, which were released at the same time. But nearly six years later, it seems that AMD are finally getting their act together and improving the experience for RDNA 2 owners. Let's see how it all works after this.
Today's video is brought to you by the AOC Aegon Pro AG276QSG2, a brilliant 1440p 360 Hz gaming monitor featuring Nvidia G-Sync Pulsar technology. You will not find a better motion experience today. G-Sync Pulsar combines the elite clarity of strobing with variable refresh rates, ensuring you get clear, fast-moving visuals no matter the situation. I tested G-Sync Pulsar earlier this year and was super impressed with how easy it is to use, how flawlessly it works, and how it's able to beat even the fastest OLEDs of today in motion clarity. The AG276QSG2 also includes an adaptive light sensor which can automatically adjust the screen to match your environment, a neat built-in headphone holder, and a fully adjustable esports stand. Gamers that demand a fast display will love G-Sync Pulsar. So, to learn more and grab an AG276QSG2 today, check the links in the description.
Okay, so the first thing I want to dive into is performance. This is, I guess, a mild concern with FSR 4.1 running on RDNA 2 graphics cards, because back when I tested the leaked int8 version last year, performance was notably lower on RDNA 2 than RDNA 3. So, let's dive into how the final FSR 4.1 release performs. Here I'm starting with Cyberpunk 2077, and I've got three graphics cards in this chart. The Radeon RX 9070 XT is running the full-fat FSR 4.1 FP8 version, showing the performance improvement from upscaling as a percentage relative to native TAA rendering. Then there's the RX 7900 XTX running the FSR 4.1 int8 version that AMD released in June. And of course, I've now also added in the RX 6950 XT running the latest FSR 4.1 release for RDNA 2 GPUs. All GPUs here are running the game at 4K with the ray tracing ultra preset for the 9070 XT and 7900 XTX, and the ultra preset with RT reflections for the 6950 XT. The reason for using different settings is to maintain a similar base frame rate, so I can more accurately compare the performance improvements of each FSR version across GPUs. The 6950 XT is slower than the other two cards, so is less capable at 4K gaming.
The 9070 XT benefits the most from FSR 4.1 FP8 upscaling. The quality mode delivers an 86% FPS improvement in this title versus native rendering. The balanced mode ran 124% faster, and the performance mode a whopping 176% faster. FSR 4.1 int8 on RDNA 3 didn't perform quite as well, but was in a similar ballpark with an 82% gain using quality upscaling. FSR 4.1 on the RX 6950 XT was a bit slower: a 72% gain using quality, 99% gain using balanced, and 137% gain using performance. This is notably lower than the upscaling performance of FSR 3.1, indicating that the FSR 4 int8 model has higher overhead, which is as expected as it's a much more complex algorithm. FSR 3.1 on the 6950 XT delivers a similar performance boost to FSR 4.1 on the 9070 XT, while FSR 4.1 on the 6950 XT is about half a tier slower. By that I mean the performance you get using FSR 4.1 balanced is halfway between the performance you get using FSR 3.1 quality and balanced. However, there is still a huge uplift to be had in titles like Cyberpunk, where the game using these settings only runs at around 20 FPS natively.
Next up is Mafia: The Old Country. In this game, I'm using the epic preset at 4K for the RDNA 4 and 3 GPUs, while the high preset is used for RDNA 2. The 9070 XT using the full version of FSR 4.1 delivers a 68% performance boost using quality upscaling, but I only saw a 52% increase using the FSR 4.1 versions for RDNA 3 and RDNA 2. All three GPUs had a similar base frame rate in this test. The 6950 XT using FSR 3.1 delivers about the same uplift as FSR 4.1 on RDNA 4 again. So, the newer architecture is clearly more efficient at processing this type of AI algorithm. On RDNA 4, you don't see much of a performance difference between FSR 4.1 and FSR 3.1, whereas FSR 4.1 is about a quality tier slower on RDNA 2 than FSR 3.1. Here, FSR 4.1 balanced is slightly slower than FSR 3.1 quality. While FSR 4.1 is a heavier algorithm on RDNA 2, it still delivers significant performance increases across all settings, as high as 85% in this test using performance upscaling. This takes the game that renders natively at 38 FPS and transforms it into 70 FPS, which is a significantly better experience. Sure, you can get 82 FPS using FSR 3.1 performance mode or 73 FPS using the balanced mode, but as we'll see in a moment, there's clear benefits to using FSR 4.1 instead.
Lastly, I have Horizon Zero Dawn Remastered. This game is included because it runs at a higher frame rate and benefits less from upscaling, so it presents a worst case of sorts. I'm using very high settings for RDNA 4 and 3, and medium settings for RDNA 2 in this test. The 9070 XT using FSR 4.1 is able to deliver an okay to good performance improvement depending on the settings used: 23% using the quality mode and 41% using the performance mode. That's similar to the FSR 3.1 results on the RX 6950 XT. However, FSR 4.1 is unable to deliver a huge gain on both the RDNA 3 and RDNA 2 GPUs. It is capable of improving performance, but in this situation by only 10% using quality upscaling and 15 to 17% using balanced upscaling. In the worst cases, like we see in this game, FSR 4.1 using the performance mode is only able to match the performance output of FSR 3.1 using the quality mode. Quite a substantial difference in render resolution there. FSR 4.1 is rendering at 1080p while FSR 3.1 can render at 1440p for the same output performance. But even still, I think most gamers would prefer to utilize FSR 4.1 for the image quality benefits.
Obviously, I haven't tested a lot of games here, but the performance trend appears to be that FSR 4.1 for RDNA 2 is quite similar to FSR 4.1 for RDNA 3 in how it runs on the respective GPU architectures. Both iterations of the upscaler for older architectures run less efficiently than FSR 4.1 FP8 on RDNA 4, but is still capable of decent performance improvements, especially if the starting frame rate is quite low. As far as I've seen, this new iteration with an optimized AI model doesn't also run substantially worse than FSR 4 int8 on RDNA 3. It's clearly been optimized for the capabilities of RDNA 2 in a way that I think is a success.
Despite FSR 4.1 running a little worse than FSR 3.1 on RDNA 2 hardware, most gamers will want to use FSR 4.1 in all situations. The simple reason for that is significantly better image quality. In this video, I've used a 1440p output resolution with balanced upscaling for all comparisons. A really tricky but still realistic configuration for that resolution. And I think 1440p is most relevant for RDNA 2 owners given the performance of those GPUs. The difference here between FSR 3.1 and FSR 4.1 is night and day. The image quality FSR 3.1 produces at 1440p using any of the settings is dreadful. When you put FSR 3.1 directly beside FSR 4.1 using the same mode, FSR 3.1 doesn't even look like it's running at the correct resolution. In comparison, it just doesn't look like 1440p. It's blurrier, it has tons of artifacts in motion, and it just looks terrible. FSR 4.1 at least looks like it's the correct resolution with the level of sharpness you'd expect from 1440p gaming, and indeed, it's much closer to native. I tested a range of games in a variety of different scenes, and not once did I think FSR 3.1 looked better. In fact, not once did I think FSR 3.1 looked even close to FSR 4.1. It is truly miles behind at this resolution. It's so obvious in some of these scenes that I don't even need to zoom in or slow down the footage. FSR 4.1 is clearly better by a mile, and this is why it's such a huge upgrade for those with RDNA 2 GPUs.
FSR 4.1 is not perfect, but it does clean up or significantly improve all the issues with FSR 3.1. The old version of FSR is disgustingly grainy in some scenes, especially anything involving foliage, where the output is so artifact-ridden it's effectively broken. FSR 4.1 is much cleaner. It's relatively grain-free and delivers an image that actually looks like foliage instead of a blurred jumble of pixels. If I didn't tell you what the resolution of these images were ahead of time, I don't think you would have guessed the FSR 3.1 image is 1440p. FSR 4.1 also cleans up TAA blur, so the sharpness of the image is similar whether you are stationary or moving. FSR 3.1 can be blurry when stationary, but becomes ultra blurry in motion, and the difference between standing still and moving can be jarring. This gives FSR 4.1 a clarity advantage in all situations, but especially in the situations that actually matter for gaming, which is when you're moving around, exploring the environment, or getting into battles. And the lower the resolution, the bigger this clarity gain is. FSR 4.1 is a significant improvement at 4K, but a monumental improvement at 1440p.
On top of this, FSR 4.1 is more stable in its output, has better fine detail reconstruction, is better equipped to handle disocclusion, is far superior at hair rendering, and generally produces lower amounts of ghosting as well. Though there were a few cases where FSR 4.1 ghosted more than the old version. In these cases, FSR 3.1 may have less ghosting, but produces worse image quality overall and is less stable in other ways.
The quality improvement is so large it consigns FSR 3.1 to the trash. Once you use FSR 4.1, you'll never want to go back to the old version of FSR upscaling, and the performance difference between the versions becomes basically irrelevant. Based on what I've seen, you're always better off using FSR 4.1 at a lower render resolution than using FSR 3.1 if you want to target a certain performance output. Even though FSR 3.1 runs faster, this can be compensated using FSR 4.1 at a lower setting for better image quality and an overall better experience. This is especially useful to know for those planning to use FSR 4.1 on a performance-constrained mobile device or a slower GPU. FSR 4.1 on the performance mode looks better than FSR 3.1 on the quality mode. The only time I'd consider touching FSR 3.1 is in dire circumstances where you need to squeeze every last drop to try and hit 30 FPS. With that said, handhelds are set to use a different, smaller FSR 4.1 model than what I've shown here, so there may be quality differences with that version.
What about this optimized FSR 4.1 model for RDNA 2 compared to the other versions of FSR 4.1 that already exist? Well, there's good news here. The new model looks virtually identical to both FSR 4.1 FP8 and the existing version of FSR 4.1 int8. I am a very hard marker with these things. It is very difficult to slip any image quality changes or reductions past me. But in most of the scenes I examined, I found it impossible to spot a difference. Even at 1440p using the balanced mode, as you're seeing here, I tried very hard to find any differences in the games I captured, but the way this model looks is the same as FSR 4.1 FP8. Every upscaler has a certain look. They have unique quirks. Each one handles things a bit differently, and once you know the patterns, you'll see similar strengths and similar artifacts crop up in similar situations, even if the game itself is different. DLSS 4.5, for example, has a distinctly different look to FSR 4.1 and to DLSS 4.0. All three versions of FSR 4.1 have the same look, generally act the same way, and produce the same sorts of artifacts in the same situations.
In scenes where I saw slight differences between FSR 4.1 FP8 and FSR 4.1 int8, I'm not confident this is actually due to the downgraded model, or whether it's subtle differences between each capture, like standing in a slightly different spot or a small timing difference between frames. If the differences I'm talking about are margin-of-error type stuff, that's a fantastic result because it means for all practical purposes the various versions of FSR 4.1 deliver the same output. At worst, maybe FSR 4.1 int8 is slightly less stable and slightly worse at handling fine detail than the full FP8 version. There's a few edge cases I saw where this is possibly the case. I don't think most gamers will notice though. 99% of the time the versions are the same. In fact, it's remarkable how little image quality has been sacrificed to optimize FSR 4.1 for RDNA 2, because FSR 4.1 int8 looks basically the same as the full version of FSR 4.1.
If you want a detailed breakdown of FSR 4.1 versus DLSS 4.5, check out the previous content I've made. I don't think it's worth going over all of that again. It'll be the same conclusion. And in those dedicated videos, I spend a lot more time diving into everything with plenty of comparisons. The release of FSR 4.1 for the RDNA 2 architecture is a massive win for owners of RX 6000 series GPUs, and it's a massive win for owners of RDNA 3.5 mobile and handheld chips as well. You could pretty much copy and paste my thoughts from the FSR 4.1 on RDNA 3 release into this video because everything I said there applies here. This release completely transforms the upscaling experience on RDNA 2 GPUs, breathing new life into products that are nearly six years old. It's very impressive to see how well AMD has managed to optimize FSR 4.1 for an old architecture that lacks a lot of modern AI processing capabilities. This FSR 4.1 int8 model looks basically identical to other versions of FSR 4.1, and it performs similarly to FSR 4.1 int8 on RDNA 3 GPUs. It's not quite as efficient as FSR 4.1 on the latest RDNA 4 architecture, but it's good enough to still deliver decent performance improvements while upscaling.
While FSR 4.1 runs a little slower than FSR 3.1, there is no reason to use FSR 3.1 anymore. FSR 4.1 absolutely smokes FSR 3.1 in image quality, especially at resolutions RDNA 2 owners realistically use, like 1440p. FSR 3.1 is diabolically bad at low resolutions. So, the upgrade is enormous visually, and even when normalizing output performance, FSR 4.1 is well ahead. In late 2026, this release is obviously great news for those still using RDNA 2 graphics, but I do have mixed feelings about the overall situation. It's nice we're getting FSR 4.1 now, something that didn't seem possible when AMD first released FSR 4 in early 2025, but at the same time, AMD took far too long to get this feature working on RDNA 2 GPUs. In that respect, the release is a bit of a fail. I feel like RX 6000 GPU owners have been burned several times here. Like with the FSR 4.1 RDNA 3 release, we now know that high-quality upscaling is actually possible on the RDNA 2 architecture.
However, buyers of RDNA 2 GPUs had to sit there watching throughout the entire DLSS generation as GeForce RTX owners got access to progressively better and better upscalers, while AMD was stuck on crappy FSR 1, then FSR 2, then FSR 3. A gamer that bought an RTX 30 series card instead of an RX 6000 card got five years of high-quality upscaling before AMD got their act together, as DLSS 2.0 really started to take off way back in 2021. Theoretically, AMD could have been much more competitive with Nvidia in upscaler quality throughout this time period, but they didn't invest anywhere near enough into their software features. RX 6000 owners were burned again when it took AMD over a year to bring FSR 4 technology to RDNA 2 officially, considering a version based on AMD source code was made available and confirmed to work in late 2025. Not to mention that FSR 4 for RDNA 4 is over 18 months old now. I know the final version is more optimized than the work-in-progress int8 version from 2025, but that's a long time to not offer anything to owners of old graphics cards. Even a beta version or something like that would have been nice.
Throughout this saga, it doesn't seem like AMD were initially willing to support their previous generation GPUs properly. I think to some degree they were dragged and pushed into it, helped along the way by Sony, who clearly wanted better technology for their PlayStation consoles that used older architectures. And if you're not aware, FSR 4.1 and PSSR used for the PS5 Pro are pretty similar. We've gotten there in the end, but it was a bit of a struggle that I think has turned some Radeon owners away from the brand. They need to do a better job of supporting existing customers moving forward. It's really that simple.
Anyway, the good news is that FSR 4.1 works well on RDNA 2 graphics cards. It already works well on newer architectures, and this will hopefully help to extend the life of aging models that are still being used today. So, anyway, that's it for this one. If you do want to support Hardware Unboxed and the independent testing that we do, the best way to do that is via Patreon. Links to those in the description below. If you sign up, you get access to all the usual benefits like our Discord community, monthly live streams, and BTS content. I am going to dive in and also look at FSR frame generation and the frame pacing improvements that have supposedly been made. So, that'll be up on the channel soon. Stay tuned for that. Subscribe if you want to. But for now, that's it for this one. Thanks for watching, and I'll catch you in the next one.