AMD has finally released FSR “Redstone,” the bundle of new machine-learning techniques it first teased at Computex. Instead of replacing FSR 4, Redstone expands it with three significant additions: ML-based frame generation, a neural denoiser for ray tracing (Ray Regeneration), and a new real-time lighting system called Radiance Caching. All three rely on the AI hardware in RDNA 4 GPUs, meaning they’re exclusive to the latest Radeon RX 9000-series.
Redstone is AMD’s clearest move yet to close the quality gap with NVIDIA’s DLSS 4. In practice, it succeeds in some areas right away, shows promise in others, and still leaves older Radeon users waiting on the sidelines.
AMD’s earlier frame-generation system in FSR 3 relied on simpler interpolation and often trailed behind DLSS in clarity. Redstone changes that. The new model combines optical-flow analysis with motion-vector re-projection and uses two neural networks: one that drives the flow re-projection and another that predicts colour during the blend stage.
The result: noticeably cleaner generated frames. In games like F1 25, shadow edges stay together rather than breaking apart, and fast motion looks more stable. Switching to the ML model costs only about 1.5% performance, a tiny tradeoff.
Redstone’s framegen can be built into the game or forced on through Radeon Software. At launch, more than 30 titles—Cyberpunk 2077, Hogwarts Legacy, God of War Ragnarök, THE FINALS, and others—support the driver-level override on RDNA 4 cards.
NVIDIA still leads in raw speed with DLSS x4 mode, but AMD has clearly caught up in quality. The only missing piece is support for higher multipliers, which feels overdue.
Ray-traced scenes often look noisy because the GPU fires only a handful of rays per pixel. A denoiser smooths this noise out. NVIDIA introduced an ML approach last year with DLSS 3.5 Ray Reconstruction. AMD’s answer is Ray Regeneration.
This new denoiser takes in rough ray-traced data—plus depth, normals, specular values, radiance, and visibility—and cleans it before upscaling or frame generation happens. It doesn’t depend on the upscaler the way NVIDIA’s method does; it can run on its own.
In Call of Duty: Black Ops 7, the first game to support it, reflections look more natural and stable. The jump isn’t dramatic in a single screenshot, but it adds realism when scenes are in motion. The performance hit is around 2%, which is about as close to “free” as these features get.
The bigger problem is availability: so far, it’s only in one game.
Radiance Caching is the forward-looking piece of Redstone, and also the hardest to judge. AMD calls it a neural radiance cache, meant to reduce the need for expensive ray-traced lighting calculations. Instead of tracing every ray to completion, the renderer can stop early and ask the cache for a lighting estimate.
The critical detail is that the cache isn’t pre-trained. It starts fresh every time a game runs and learns continuously using live ray-tracing samples. That puts it closer to a real-time lighting algorithm than a traditional caching system. It also explains why it needs RDNA 4 hardware: the model trains and infers every frame.
This learning process isn’t free. The documentation notes that flicker or instability can appear in harsh lighting conditions, and developers will need to tune how fast the model learns and how its output is smoothed. AMD recommends pairing it with more innovative sampling systems, such as ReSTIR, to feed the cache with cleaner data.
Radiance Caching is available to developers now as a technical preview. Games aren’t expected to ship with it until 2026, with Warhammer: Darktide among the first.
Redstone’s ML features run only on RDNA 4 GPUs. Older cards can fall back to shader-based versions of upscaling and frame generation. Still, they don’t get the improved quality or the real ML models. If you’re on RDNA 2 or 3, nothing in Redstone changes your experience meaningfully.
One area where AMD does deserve clear credit is software polish. Radeon Settings can now override older in-game FSR versions automatically—no DLL swapping, no waiting for a patch, and no guesswork. The overlay shows exactly which Redstone features are active. This is a simple, well-executed idea that makes a real difference.
FSR 4 Redstone is the most convincing step AMD has taken toward matching NVIDIA’s image-quality tech.
• ML Frame Generation is finally competitive with DLSS and sometimes looks better.
• Ray Regeneration improves stability and clarity with almost no performance penalty.
• Radiance Caching points to a brighter future for real-time lighting.
• Driver-level overrides significantly improve how FSR is delivered to players.
• Only RDNA 4 users benefit. There’s still no official ML upscaler for older GPUs.
• Ray Regeneration support is limited to one game at launch.
• Radiance Caching is years away from real adoption.
• No x3 or x4 framegen modes yet.
If you own an RX 9000-series GPU, Redstone is a meaningful upgrade. If you don’t, it’s mostly a preview of where AMD’s rendering tech is going. The company is clearly investing in machine learning across the pipeline, and the early results are encouraging. Now it needs game and broader hardware support before Redstone becomes a valid default option rather than a promising set of new tools.
Redstone is AMD’s clearest move yet to close the quality gap with NVIDIA’s DLSS 4. In practice, it succeeds in some areas right away, shows promise in others, and still leaves older Radeon users waiting on the sidelines.
ML Frame Generation: The Big Upgrade
AMD’s earlier frame-generation system in FSR 3 relied on simpler interpolation and often trailed behind DLSS in clarity. Redstone changes that. The new model combines optical-flow analysis with motion-vector re-projection and uses two neural networks: one that drives the flow re-projection and another that predicts colour during the blend stage.
The result: noticeably cleaner generated frames. In games like F1 25, shadow edges stay together rather than breaking apart, and fast motion looks more stable. Switching to the ML model costs only about 1.5% performance, a tiny tradeoff.
Redstone’s framegen can be built into the game or forced on through Radeon Software. At launch, more than 30 titles—Cyberpunk 2077, Hogwarts Legacy, God of War Ragnarök, THE FINALS, and others—support the driver-level override on RDNA 4 cards.
NVIDIA still leads in raw speed with DLSS x4 mode, but AMD has clearly caught up in quality. The only missing piece is support for higher multipliers, which feels overdue.
Ray Regeneration: Small Hit, Clear Gains
Ray-traced scenes often look noisy because the GPU fires only a handful of rays per pixel. A denoiser smooths this noise out. NVIDIA introduced an ML approach last year with DLSS 3.5 Ray Reconstruction. AMD’s answer is Ray Regeneration.
This new denoiser takes in rough ray-traced data—plus depth, normals, specular values, radiance, and visibility—and cleans it before upscaling or frame generation happens. It doesn’t depend on the upscaler the way NVIDIA’s method does; it can run on its own.
In Call of Duty: Black Ops 7, the first game to support it, reflections look more natural and stable. The jump isn’t dramatic in a single screenshot, but it adds realism when scenes are in motion. The performance hit is around 2%, which is about as close to “free” as these features get.
The bigger problem is availability: so far, it’s only in one game.
Radiance Caching: The Most Ambitious Part
Radiance Caching is the forward-looking piece of Redstone, and also the hardest to judge. AMD calls it a neural radiance cache, meant to reduce the need for expensive ray-traced lighting calculations. Instead of tracing every ray to completion, the renderer can stop early and ask the cache for a lighting estimate.
The critical detail is that the cache isn’t pre-trained. It starts fresh every time a game runs and learns continuously using live ray-tracing samples. That puts it closer to a real-time lighting algorithm than a traditional caching system. It also explains why it needs RDNA 4 hardware: the model trains and infers every frame.
This learning process isn’t free. The documentation notes that flicker or instability can appear in harsh lighting conditions, and developers will need to tune how fast the model learns and how its output is smoothed. AMD recommends pairing it with more innovative sampling systems, such as ReSTIR, to feed the cache with cleaner data.
Radiance Caching is available to developers now as a technical preview. Games aren’t expected to ship with it until 2026, with Warhammer: Darktide among the first.
Platform Support and Driver Integration
Redstone’s ML features run only on RDNA 4 GPUs. Older cards can fall back to shader-based versions of upscaling and frame generation. Still, they don’t get the improved quality or the real ML models. If you’re on RDNA 2 or 3, nothing in Redstone changes your experience meaningfully.
One area where AMD does deserve clear credit is software polish. Radeon Settings can now override older in-game FSR versions automatically—no DLL swapping, no waiting for a patch, and no guesswork. The overlay shows exactly which Redstone features are active. This is a simple, well-executed idea that makes a real difference.
Verdict: Strong Progress, Not a Complete Package
FSR 4 Redstone is the most convincing step AMD has taken toward matching NVIDIA’s image-quality tech.
What works well:
• ML Frame Generation is finally competitive with DLSS and sometimes looks better.
• Ray Regeneration improves stability and clarity with almost no performance penalty.
• Radiance Caching points to a brighter future for real-time lighting.
• Driver-level overrides significantly improve how FSR is delivered to players.
Where Redstone still falls short:
• Only RDNA 4 users benefit. There’s still no official ML upscaler for older GPUs.
• Ray Regeneration support is limited to one game at launch.
• Radiance Caching is years away from real adoption.
• No x3 or x4 framegen modes yet.
If you own an RX 9000-series GPU, Redstone is a meaningful upgrade. If you don’t, it’s mostly a preview of where AMD’s rendering tech is going. The company is clearly investing in machine learning across the pipeline, and the early results are encouraging. Now it needs game and broader hardware support before Redstone becomes a valid default option rather than a promising set of new tools.