AMD is working on a version of FSR 4 for older Radeon graphics cards, and today we are testing a work-in-progress build to see how it's shaping up. Is this version as good as FSR 4 in terms of image quality? How does it perform on RDNA 3 and RDNA 2 GPUs? And, on balance, is it going to be a viable option for upscaling if AMD releases it officially? We are going to answer all of those questions today.
As we know, FSR 4 is only officially supported on the latest RDNA 4 graphics cards, such as the Radeon RX 9070 XT. This is because FSR 4 uses the FP8 data format to accelerate the AI upscaling algorithm, which is only supported in AMD's RDNA 4 architecture.
However, AMD has been developing a version of FSR 4 that uses the INT8 format instead, which is more broadly supported. AMD introduced INT8 support with the RDNA 2 architecture, so theoretically, the INT8 version of FSR 4 could work on GPUs as old as the RX 6000 series.
The only reason we know AMD is developing an INT8 version of FSR 4 is that the company briefly published source code for it on its GPUOpen GitHub repository. People quickly downloaded the INT8 model library before AMD removed it, and then compiled that library into a working FidelityFX DLL that can be inserted into games. This is how we are able to test the INT8 version before AMD makes it official – it has been a community effort over the last month to get it working.
How to Get FSR 4 Working on RDNA 3
Getting FSR 4 working on RDNA 3 GPUs is surprisingly easy once you've downloaded the INT8 DLL file. All you need to do is find a game that already supports FSR 3.1 and copy the DLL into the game's directory, overwriting the existing FidelityFX DLL.
This replaces FSR 3.1 with FSR 4 in the game, and depending on how the game implemented FSR 3.1 support, the settings menu may even show FSR 4 instead of FSR 3.1. Getting FSR 4 running on RDNA 2 is a bit trickier due to some driver issues, but we used the guide from Ancient Gameplays to get it working.
Before we get into quality comparisons and performance numbers, a quick disclaimer: this testing involves what appears to be an in-development version of INT8 FSR 4, so we don't know if it will be representative of the final version released to the public. AMD could optimize it further before launch, but this still gives us a solid look at how it's developing so far.
FSR 4 on RDNA 4 vs. FSR 4 on RDNA 3
The first thing we wanted to examine is whether the FSR 4 INT8 model is the same as the FSR 4 FP8 model and whether it delivers the same image quality. Our guess is that it's not identical – the model has likely been modified or scaled back to use INT8 instructions. Let's take a closer look.
For a better representation of image quality comparisons, check out the HUB video below:
Here, we are comparing the FSR 4 INT8 version running on a Radeon RX 7900 XTX to our FSR 4 FP8 captures from our initial analysis a few months ago, which ran on the RX 9070 XT. At high render resolutions like 4K Quality mode, the INT8 and FP8 versions are remarkably similar in visual quality, to the point where most people would find them virtually indistinguishable.
There are some minor differences, so the two models are not exactly the same, but the INT8 version – at least in this configuration – is a very good proxy for the standard FSR 4 we analyzed earlier in the year.
It's clear from examining the INT8 version closely that it uses the same type of algorithm as FSR 4 FP8. In other words, the general output, strengths, and weaknesses of both versions are similar. Both can reduce or eliminate TAA-related blur in motion, handle disocclusion effectively, and significantly improve over FSR 3 in areas like shimmering and transparencies. This does not appear to be a case where the INT8 version is a completely reworked technology merely labeled "FSR 4."
The main difference between the two models lies in how they handle detail reconstruction, especially at lower render resolutions. The FP8 model is better at determining which pixels should form the final image and which elements are actually being rendered. This usually isn't a problem at 4K in Quality mode because the INT8 model has plenty of data to work with, so we only noticed minor differences – for example, when transparent objects overlapped, the FP8 model provided better definition and less blur.
The FP8 model delivers progressively better image quality than the INT8 model as render and output resolutions decrease. At 4K in Performance mode, the two models still look similar, but the FP8 model's output is slightly better than what we saw when comparing the Quality modes.
For a better representation of image quality comparisons, check out the HUB video below:
At 1440p Quality mode, we start to see more noticeable differences, including a generally sharper presentation with the FP8 version. At 1440p Performance mode, it becomes quite clear that the original FP8 image is less blurry, more detailed, and sharper overall.
In the configurations most gamers would likely use, such as 4K Quality or 1440p Quality, the INT8 and FP8 versions of FSR 4 are very close
That's not to say the INT8 version performs poorly at low render resolutions such as 1440p Performance mode – it still holds up well compared to FSR 3 – but it's clear that this model isn't as capable of handling low-resolution upscaling as the full FSR 4. The AI model size has likely been reduced for performance reasons, which limits its effectiveness in these scenarios. We didn't test 1080p upscaling, but we suspect those configurations would show a continuation of this trend.
The good news is that this reduction in detail and sharpness seems to be the main difference in image quality. At 1440p, the INT8 version is not noticeably less stable – the levels of shimmering and aliasing are comparable between the INT8 and FP8 versions. The INT8 version is simply less sharp and detailed.
This is actually a better outcome than the reverse situation, where sharpness is preserved but stability is lost, as that kind of artifact is highly noticeable and one of the main reasons FSR 3 looked poor at 1440p.
To some extent, comparing the versions at 1440p in Performance mode is nitpicking, since we don't recommend using that mode even for the full FSR 4 or DLSS 4. In the configurations most gamers would likely use, such as 4K Quality or 1440p Quality, the INT8 and FP8 versions of FSR 4 are very close – and at the highest resolutions, even closer than one might expect. We suspect that many people would not notice the differences in some of the examples shown here so far.
Performance Benchmarks
The next part of the puzzle is performance. The FSR 4 INT8 model is more demanding to run on RDNA 3 and RDNA 2 GPUs than the FSR 4 FP8 model is on RDNA 4. Let's go through some performance numbers to illustrate what we mean.
In Ratchet and Clank: Rift Apart at 4K on the RX 9070 XT, FSR 4 and FSR 3 offer somewhat similar performance. At worst, FSR 4 is about half a quality tier slower. For example, the performance uplift you get from FSR 4 Balanced mode sits between the uplifts of FSR 3 Quality and Balanced modes. In other words, it delivers a slightly smaller boost at the same quality setting. This serves as our baseline for the full version of FSR 4.
When testing FSR 4 INT8 on the RX 7900 XTX at 4K in the same game, we see a more noticeable performance impact. In this configuration, FSR 4 is roughly one full tier slower. The uplift from Balanced mode is similar to what FSR 3 achieves in Quality mode. We also included XeSS DP4a results since many RDNA 3 owners use that technology instead of FSR 3, and in this case, FSR 4 Balanced performs similarly to XeSS Quality.
The INT8 version performs even slower on the RX 6750 XT, an RDNA 2 card, when testing at 1440p. Here, FSR 4 Performance mode is slightly slower than FSR 3 Quality mode and roughly equivalent to XeSS Ultra Quality. This makes the algorithm quite demanding for this older mid-range GPU.
In The Last of Us Part I, the full FP8 FSR 4 on the RX 9070 XT is about one tier slower than FSR 3, meaning FSR 4 Balanced mode delivers performance similar to FSR 3 Quality. On the RX 7900 XTX, the FSR 4 INT8 version is about two tiers slower, so FSR 4 Performance mode provides an uplift comparable to FSR 3 Quality. We still measured a 23% performance increase using FSR 4 Quality mode, which isn't bad, but it falls short of the 44% increase achieved with FSR 3 Quality mode, showing how much more demanding the newer algorithm is. On the RX 6750 XT at 1440p, we saw similar results: FSR 4 INT8 Performance mode was slightly slower than FSR 3 Quality, and FSR 4 Quality mode only delivered a 15% improvement.
Horizon Zero Dawn Remastered doesn't benefit as much from upscaling as the other titles we tested. On the RX 9070 XT, FSR 4 runs about half to one full tier slower than FSR 3. On the RX 7900 XTX, performance drops roughly two tiers, meaning FSR 4 INT8 Performance mode matches FSR 3 Quality and XeSS Quality. We observed just a 10% improvement using FSR 4 Quality mode, which isn't a major gain, although image quality remains excellent.
The RX 6750 XT performs even worse here, showing roughly a three-tier difference between FSR 4 and FSR 3. FSR 4 Performance mode still trails behind FSR 3 Quality in FPS output, only matching XeSS Ultra Quality.
The last game we tested for performance was Spider-Man 2. On the RX 9070 XT, FSR 4 was about one quality tier slower than FSR 3. On the RX 7900 XTX, the FSR 4 INT8 version ran about two tiers slower, so once again, FSR 4 Performance mode matched FSR 3 Quality and XeSS Quality. The RX 6750 XT showed a similar two-tier gap at 1440p – slightly better than Horizon Zero Dawn – but using Quality mode only yielded around a 10% performance boost.
Overall, based on AMD's early published source code, the current INT8 mode of FSR 4 typically requires dropping two quality modes to achieve the same performance uplift as previous upscalers. For example, if you currently use FSR 3 Quality or XeSS Quality, you would likely need to switch to FSR 4 INT8 Performance mode to maintain similar frame rates.
Alternatively, you can choose a higher quality mode, but you'll have to accept a smaller performance increase compared to native rendering. The Quality mode may only deliver a 10 – 20% boost, whereas around 30% is common when using the full FP8 version of FSR 4 on an RDNA 4 GPU.
Image Quality Comparison
FSR 4 INT8 vs FSR 4 FP8 vs FSR 3 vs XeSS
Let's jump back to image quality comparisons to explore how these stack up. We'll start with everything set to the same labeled mode – in this case, Quality mode. Keep in mind that FSR 4 INT8 runs slower than the other techniques, but this comparison shows what to expect when using a typical level of upscaling without focusing too much on performance gains.
For a better representation of image quality comparisons, check out the HUB video below:
The ranking from best to worst image quality is clear when viewing the examples side by side. FSR 4 delivers the best results, with the FP8 mode performing slightly better than the INT8 mode, especially at higher resolutions. FSR 4 INT8 then ranks ahead of XeSS's DP4a mode, which runs on RDNA 3, and in last place is FSR 3, which provides the weakest image quality. XeSS DP4a is closer in quality to FSR 4 INT8 than to FSR 3, particularly at lower resolutions.
The main reasons FSR 4 INT8 looks better than XeSS and especially FSR 3 are its reduced motion blur, higher detail, and greater image stability.
The stability difference between FSR 4 INT8 and FSR 3 is dramatic – that's clearly the biggest improvement among these technologies. FSR 4 INT8 also outperforms XeSS in this area across most examples we examined. It is noticeably sharper than both XeSS and FSR 3. XeSS performs reasonably well, but even at 1440p Quality mode, FSR 4 INT8 maintains a clear edge.
We also tested XeSS at Ultra Quality, which uses the same render resolution as FSR in Quality mode. The difference was minimal, and we would still choose FSR 4 INT8 over XeSS at the same resolution. This confirms that the FSR 4 INT8 model is currently the best image quality option available for GPUs that do not support FSR 4 or DLSS, making it our clear recommendation for RDNA 3 and RDNA 2 graphics cards.
FSR 4 INT8 vs FSR 4 FP8 vs FSR 3 vs XeSS (Normalized)
But what if we normalize for output performance? In this case, we compare FSR 3 and XeSS using Quality mode to FSR 4 FP8 using Balanced mode and FSR 4 INT8 using Performance mode. Does the FSR 4 INT8 model still hold up?
For a better representation of image quality comparisons, check out the HUB video below:
At 4K, the answer is definitely yes. While this adjustment increases the performance gap between the FP8 and INT8 models, FSR 4 INT8 in 4K Performance mode remains a strong option.
In all tests, we preferred FSR 4 over FSR 3 in Quality mode because FSR 4 produces sharper, more detailed, and more stable images. This is a significant win for gamers who typically use FSR, especially in games that lack XeSS support. FSR 4 INT8 Performance mode outperforms FSR 3 Quality mode, which mirrors what we observed with the full FP8 version of FSR 4.
This alone explains why RDNA 3 gamers are eager to get access to FSR 4. Compared to FSR 3, players effectively have two choices: a higher-quality but slower upscaling option in FSR 4 Quality mode, or a similarly performing yet superior option in FSR 4 Performance mode. Either way, the INT8 version provides a better overall experience. We would only recommend reverting to FSR 3 if no FSR 4 mode meets your desired performance target.
In this performance-normalized comparison, FSR 4 INT8 generally outperforms XeSS, even when comparing Performance mode to XeSS Quality mode. FSR 4 exhibits less shimmering and aliasing than XeSS while maintaining better sharpness and reduced blur. The difference is closer than that between FSR 4 and FSR 3, and there are cases where XeSS produces slightly better detail in specific image areas due to its higher render resolution. However, overall, we would still choose FSR 4 at a lower mode rather than XeSS at a higher one.
For a better representation of image quality comparisons, check out the HUB video below:
Much to our surprise, the advantage of FSR 4 INT8 over other upscalers grows at 1440p. At this resolution, it is clearly the best choice for users without access to FSR 4 or DLSS. FSR 4's FP8 model remains superior, offering a sharper and clearer image in motion, but the INT8 version still performs very well. It is clearly better than FSR 3 at 1440p, even when comparing Performance to Quality modes – the improvement is dramatic. Once you experience the INT8 variant at this resolution, returning to FSR 3 is not an option.
We also found that FSR 4 INT8 typically provides better image quality than XeSS when performance-normalized. XeSS DP4a struggles in certain areas at this resolution, even in Quality mode – for example, grass exhibits noticeable shimmering and grain.
FSR 4 is a more stable upscaler, despite rendering at a lower resolution, and generally produces a similar or better level of sharpness with less blur. While XeSS occasionally delivers more fine detail, that advantage is not enough to offset its reduced stability. In actual gameplay, FSR 4 INT8 feels like it produces a higher-resolution output overall.
This is especially encouraging for users with lower-end RDNA 3 and RDNA 2 graphics cards, who are more likely to use FSR 4 INT8 at 1440p or 1080p. It should be the best option for improving performance, even if it is somewhat more demanding than other upscaling technologies.
Superior Upscaling
There is no doubt that AMD should release the INT8 version of FSR 4. The in-development build we analyzed – made possible by AMD's source code leak – is a significant upgrade over FSR 3 and an improvement on XeSS, which owners of older Radeon cards often rely on.
For those with RDNA 3 and RDNA 2 GPUs, FSR 4 INT8 would be the go-to choice for upscaling in modern titles. It would extend the life of those products and serve as an excellent way to support loyal Radeon users.
We were skeptical about whether AMD could bring FSR 4 to older graphics cards in a way that worked effectively. It seemed likely they could get it functioning at a basic level, but that the algorithm might be too demanding without the hardware acceleration capabilities introduced in RDNA 4. This is partially true – FSR 4 INT8 uses a simplified algorithm that is more taxing than FSR 4 FP8 – but despite that, it still performs as a capable and impressive upscaler.
We underestimated how poor FSR 3 looks at resolutions like 1440p and how much room there was for improvement, even if the heights of full fat FSR 4 are not reached.
In many situations, it is remarkable how close the INT8 version comes to the FP8 model in image quality. Even though it is more performance-intensive on RDNA 3, users may still be willing to trade a few frames per second for the visual upgrade it provides. If AMD offers it as an optional toggle, allowing players to switch between FSR 4 and FSR 3, there would be no real downside to making it available.
The question now is whether AMD will release it. AMD did not respond to our questions before publication, but since this is official AMD code, they could make it public at any time. One possible reason for holding it back might be to avoid cannibalizing RDNA 4 sales, as FSR 4 is one of its major selling points.
If FSR 4 became available for RDNA 2 and RDNA 3, owners of those cards might be less inclined to upgrade. However, that would be a poor decision – it would not only appear anti-consumer, but it would also risk alienating existing Radeon customers. Supporting those users is more likely to strengthen brand loyalty and encourage future upgrades. Moreover, AMD did not sell a large number of previous-gen GPUs, so the real competitive advantage of FSR 4 lies in the broader battle between Radeon and GeForce, not between new and old Radeon models.
We will continue monitoring this situation closely. If AMD officially releases the INT8 version, we will revisit it to see how it compares to the current build and assess what improvements have been made.