I put AMD’s new AI frame generation to the test to see if it lives up to the hype.

It’s been nearly seven months since its announcement, but AMD’s highly anticipated machine learning-enhanced update to FSR has finally arrived. With the release of the AMD Adrenalin 25.12.1 driver today, those with an RDNA 4 card can dive into the new FSR experience. Known as FSR ‘Redstone’, this update positions AMD to compete more closely with Nvidia and Intel in the realm of AI-driven upscaling, frame generation, and other rendering technologies.

That said, it’s important to note that AMD still lacks multi-frame generation, so it doesn’t quite match the competition yet. Additionally, it’s still early to assess whether the new FSR frame generation holds up against DLSS MFG or XeSS-MFG in terms of image quality, especially since it currently supports a limited selection of games.

Another point of confusion is AMD’s decision to drop any numbering from FSR. For instance, FSR 4 was introduced in February, bringing AI-powered upscaling to RDNA 4 graphics cards, but it’s now simply referred to as AMD FSR Upscaling. This name overlaps with FSR 3’s upscaler and FSR 2, which do not utilize machine learning. To clarify, AMD has introduced more specific terms: AMD FSR Upscaling (Analytical) and AMD FSR Upscaling (ML).

The same renaming applies to FSR 3 Frame Generation, which is now called AMD FSR Frame Generation (Analytical), while the new Redstone version is labeled with ML in parentheses. This naming convention is a bit frustrating, as it complicates understanding the differences between the various iterations.

FSR Redstone brings three features, although one—FSR Ray Regeneration—has already been released. The remaining two, FSR Frame Generation and FSR Radiance Cache, are not available for preview, with the latter set to debut in games next year. Therefore, my testing of FSR Redstone focuses solely on the new frame generation system.

Fortunately, it performs impressively. The AI-driven color calculations for generated frames significantly reduce the odd artifacts seen with FSR 3 frame generation, making for a smoother experience.

Currently, FSR Frame Generation (the ML version) is compatible with only 31 games. I selected a few to test, first using FSR 3 upscaling and frame generation, then switching to Redstone.

My test rig featured a Ryzen 9 9900X on an MSI MEG X870 Godlike motherboard, with 32 GB of DDR5-6000 CL32 RAM and an Asus Prime Radeon RX 9070 XT. For some titles, I utilized built-in benchmarks to provide a direct comparison of visuals and performance.

To enable Redstone (FSR 4 upscaling and the new AI frame generation), you first activate FSR 3 or FSR 3.1 in the game, then exit. In Adrenalin Software, you toggle the relevant settings for the game, forcing the drivers to use the desired FSR version.

Until developers implement Redstone directly, this is the only method available. Some games, like Cyberpunk 2077, automatically convert FSR 3/3.1 upscaling to FSR 4, but it may take time for consistent naming conventions to be adopted across titles.

I tested Black Myth: Wukong at 4K using the Cinematic graphics preset, without Full Ray Tracing enabled. For both captures, I used Performance upscaling (50%) along with frame generation. I also neglected to move the cursor out of the frame, which I only noticed too late—my apologies for that.

While only capturing half the frame isn’t ideal for evaluation, it’s clear that FSR Redstone outperforms FSR 3. Most of the quality improvements stem from FSR 4—now referred to as FSR Upscaling (ML)—which preserves nearly all native texture details and avoids blurring objects like falling leaves.

The frame generation is well-executed in this game, making it challenging to discern the improvements in the full videos. However, during fast camera movements, the differences become apparent: FSR 3 frame generation can blur and distort objects, while the AI system maintains clarity.

Best of all, there’s no performance drop when using Redstone frame generation compared to FSR 3. While FSR 4 does incur a minor performance hit, it’s relatively insignificant.

In Cyberpunk 2077, tested again at 4K with the RT Ultra graphics preset, there’s a slightly larger performance impact with Redstone than in Black Myth: Wukong, but it’s still manageable, and the upscaler is responsible for this, not the frame generator.

The improvements in detail retention for dust, fog, and smoke with AMD’s machine learning are noticeable compared to standard FSR 3. However, like Wukong, the older technology is well-implemented, so the enhancements are mostly subtle.

Redstone’s frame generation handles rapid camera movements and distant thin objects, like power lines, quite well, though I did notice some shimmering on certain surfaces during gameplay. I attempted to capture these artifacts, but they often vanished in video compression.

One game that could greatly benefit from AI enhancements is F1 25. The first 30 seconds of the video clearly illustrate the shortcomings of FSR 3 frame generation, particularly in how it handles shadows on cars.

In contrast, Redstone generates shadows much more accurately, although some glitches remain with distant cars. This issue likely stems from the engine using low-resolution shadows for vehicles beyond a certain distance from the camera.

Once again, all of this is achieved with only a minimal performance impact compared to FSR 3. Without upscaling, F1 25 runs at around 40 frames per second at 4K Ultra High, with 1% lows of 29 fps. If you have an RDNA 4 card, head into Adrenalin Software, toggle the settings, and launch the game. Once in, activate FSR 3 upscaling and frame generation, and the software will take care of the rest.

For a final test of FSR Redstone’s frame generator, I chose GTA 5: Enhanced from the list of 31 games. I initially attempted to test Mafia: The Old Country, but its settings menu was completely blank for me, no matter what I tried.

Testing GTA 5 can be tricky, as restarting the game means you don’t always return to the exact save point. So, please excuse the lack of synchronization in the videos, and focus on comparing how FSR 3 performs against the AI enhancements.

Visually, there isn’t a significant difference. Redstone does a slightly better job of keeping power lines sharp and preventing foliage from blurring during rapid movement, but the contrast isn’t stark.

At 4K with the High RT graphics preset, GTA 5 showed the largest performance drop of all the games tested, averaging around 30 fps lower and 25 fps lower for 1% lows. The ray tracing in this game can be demanding on a GPU, especially at high settings, which likely contributes to the performance challenges when combining ray tracing and AI upscaling.

While Call of Duty: Black Ops 7 is one of the few games that supports the AI frame generation, I focused instead on FSR Ray Regeneration. This feature aims to improve denoising and reconstructing images from reflected surfaces, similar to Nvidia’s DLSS Ray Reconstruction.

Black Ops 7 is currently the only game that supports this feature, which is an unusual choice for showcasing the technology. After all, who’s going to scrutinize the accuracy of reflections on a shiny floor during an intense battle?

The video compares standard Redstone (FSR 4 + AI frame generation) with Redstone featuring Ray Reconstruction. If you look closely, you can see where the AI denoiser enhances the visuals. Interestingly, the frame generation also introduces some issues, though they’re not easily noticeable.

Between 19 and 20 seconds into the video, the standard Redstone run generates some odd artifacts, but these are only visible when scrutinizing frame by frame, so they wouldn’t be apparent during actual gameplay. This highlights that, while Redstone is impressive, it’s not without its flaws.

FSR Ray Regeneration doesn’t seem to impact performance significantly in Black Ops 7, at least not according to the game’s benchmark. We’ll need more extensive testing to form a comprehensive judgment on its effectiveness.

The final aspect of FSR Redstone is a bit peculiar, as it’s not something PC gamers can directly utilize; it must be implemented by developers within their engines. This applies to all parts of FSR, as you can’t force it to work in a game that hasn’t integrated it, but there is some user interactivity with the other features.

AMD’s Josh Hort, Senior Director and Head of ISV Enabling, explained the process behind FSR Radiance Cache: “We sample the scene, the camera viewport and geometry are used to generate accurate radiance values, which are then fed to a neural network for training. We train the network to understand how radiance behaves in the scene, how light interacts with surfaces, how indirect lighting propagates, and so on.

“During runtime, in the FSR radiance caching phase, we use the trained network to provide radiance results early, as soon as the second ray intersection. This allows us to achieve high-quality lighting much faster with fewer rays and less brute force computation, directly translating to improved performance.”

What’s still unclear is what Hort means by ‘we’: Is the training entirely handled by AMD, or does it require developer involvement? An updated FSR SDK with a preview of FSR Radiance Cache should be available on GPUopen by the time this article is published, so I hope to check it out then.

AMD showcased a brief clip of Warhammer 40,000: Darktide running with FSR Radiance Cache, and while it looked impressive, the lack of performance metrics made it difficult to assess its effectiveness, as improving ray tracing speed is the primary goal of this feature.

To sum it up, FSR Redstone is a notable success. The AI upscaler performs admirably, and the AI frame generator may be even better. However, a single sample of the AI denoiser isn’t enough to determine if it’s always worth enabling, but initial impressions are positive.

That said, AMD isn’t off the hook. They are quite late to the market; Nvidia introduced AI-powered frame generation back in September 2022, and Intel followed in December 2024, both releasing updated versions since.

AMD’s RDNA 4 GPUs have been available since March, and while FSR 4 was rolled out quickly, it shouldn’t have taken this long to launch FSR Frame Generation. The priority given to FSR Ray Regeneration is puzzling. It seems AMD initially intended for FSR 4 to encompass the full Redstone package, but development delays pushed it back.

Another concern is the naming convention. While the full names may not be used in games—just FSR Upscaling, etc.—AMD could have simply labeled Redstone as FSR 5 or even 4.5, or any version number.

Having a number is crucial for distinguishing between updates. When new versions inevitably arrive, how will gamers and developers differentiate them? While version codes will be used by AMD’s FSR developers, burying them under marketing jargon is counterproductive.

Lastly, I’m not thrilled that accessing the main features of the new FSR requires using the Adrenalin Software tool. This was already the case with FSR 4, but now it applies to the ML version of frame generation as well. While it’s not a major issue—if you have an RDNA 4 graphics card, you’ll want to use Redstone—it can be a bit overwhelming for newcomers to PC gaming.

The engineers behind AMD’s GPUs and software are incredibly talented, and both RDNA 4 and FSR are impressive technologies. It’s just unfortunate that the marketing side of things could use some improvement.