What if your GPU could “guess” missing light bounces, interpolate a convincing extra frame, and upscale a low-res render so well you forget it wasn’t native? That’s the pitch behind AMD’s FSR “Redstone” — not a single tool but a suite of ML-powered tricks AMD hopes will narrow the gulf with Nvidia’s DLSS and change how games get made and played.
Four pieces, one umbrella
Redstone bundles four capabilities under the FSR name: the upgraded FSR upscaler (the evolution many outlets call FSR 4), ML-driven Frame Generation, Ray Regeneration (an ML denoiser/repair step for ray-traced passes), and Radiance Caching (an ML-based system to predict indirect lighting). They’re designed to be mixed and matched by developers: use one, use all, or let drivers swap in fallbacks for older hardware.
Short version of why each matters:
- Upscaling: ML-based upscaling that runs on AMD’s new on-chip AI units, meant to beat earlier temporal FSR results.
- Frame Generation: AI interpolation that inserts frames to boost perceived frame rates while aiming to keep latency and artifacts low.
- Ray Regeneration: A neural step to clean up sparse ray-tracing samples so you get convincing reflections and lighting without tracing every ray.
- Radiance Caching: Predicts bounced light so the GPU doesn’t have to compute full multi-bounce global illumination for every pixel.
AMD’s demos and slides show big-sounding gains — multiples of native 4K performance in some titles when Redstone features are stacked — but they come with a long list of caveats that matter in practice.
Hardware and availability: gated by RDNA 4
There’s a catch: most of the ML-driven pieces require RDNA 4 hardware, i.e., Radeon RX 9000-series cards. AMD added dedicated AI hardware to RDNA 4, and Redstone’s models run there. Older RDNA GPUs can fall back to shader-based implementations, but without the same quality or performance.
Not all parts are shipping to consumers at once. Upscaling, frame generation, and Ray Regeneration are already appearing in a handful of games; Radiance Caching is available to developers through AMD’s SDK now, with game integrations expected in 2026. That staggered rollout shapes how quickly players will feel the impact.
How Redstone stacks up against DLSS (and why that comparison matters)
Nvidia has a multi-year lead in ML upscaling + frame-gen + ray-tracing reconstruction, and DLSS’s evolution — including multi-frame generation and ray reconstruction — set expectations. AMD’s Redstone is clearly chasing that functionality, but AMD has chosen to gate the ML models behind newer silicon just as Nvidia did for its advanced features.
The result is familiar: great technology on paper, but limited reach at first. Redstone closes some quality gaps, but it doesn’t leapfrog DLSS in every respect — instead it levels the playing field for AMD’s latest hardware.
Real-world trade-offs
Frame generation can multiply perceived frame rates, but it isn’t magic. It works best when a game already delivers a reasonably high base frame rate after upscaling (AMD recommends having ~60 fps post-upscale to avoid odd artifacts). There’s also the latency question: generated frames can add responsiveness trade-offs that vary game by game.
Ray Regeneration and Radiance Caching promise big win-for-cost improvements in ray-traced visuals, but they rely on smart sampling and ML denoising. If a game’s engine or the dev’s implementation is sloppy, artifacts will crop up. Conversely, done right, these techniques let developers add cinematic lighting effects without murdering performance.
Industry momentum and partnerships
AMD isn’t working alone on this. Microsoft has been a visible partner in developing Ray Regeneration, and Xbox teams have already been involved in co-development efforts — a collaboration Windows Central highlighted when Microsoft’s Matt Booty described their close work with AMD. That partnership hints at Redstone-inspired ideas moving into console hardware and future Xbox generations; Microsoft’s public remarks suggest the tech could influence next-gen boxes and first-party titles alike. For more on Microsoft’s angle, see the coverage of future Xbox work with AMD.
Sony also has prior ties with AMD around FSR work (PlayStation contributed to earlier FSR development), and observers have speculated Redstone-like features could matter for future PlayStation hardware or mid-cycle refreshes. Console implications are real: better upscaling and frame gen benefit handhelds, mid-range consoles, and power-constrained systems. For how streaming and PlayStation hardware are evolving in parallel, see the piece about the PlayStation Portal streaming update. If you’re shopping for consoles, retailers and fans are already eyeing what a PlayStation refresh or successor might gain from these techniques — the rumored PlayStation 5 Pro and successors are the obvious beneficiaries (the PlayStation 5 Pro has popped up in retail listings and speculation; you can check the PlayStation 5 Pro availability on Amazon PlayStation 5 Pro).
Developer ergonomics
AMD has made Redstone an extension of the existing FSR ecosystem so that games already using FSR 3.1 or FSR 4 can relatively easily adopt the new features. Where a game supports FSR 3.1, a DAO-style driver or simple DLL swap can let AMD’s stack plug in the newer ML models on supported hardware; older GPUs fall back to FSR 3.1 mode. Radiance Caching is being released as an SDK to developers so they can experiment now and ship integrations next year.
Adoption will still depend on how persuasive the gains look for publishers and how large the RDNA 4 install base becomes. AMD’s early partner wins (a few big PC titles shipping with parts of Redstone) are promising, but broad uptake will be a slow burn.
The pragmatic view
Redstone is less a single “win” and more a coordinated push: it packages ray-trace acceleration, denoising, frame interpolation, and ML upscaling into something game teams can reasonably bolt into modern engines. For players with RX 9000-series cards, the results will be tangible: higher frame rates, better ray-traced imagery at lower cost, and an upscaler that looks closer to native than older FSR versions.
For everyone else, the payoff is indirect for now — developers may tune games assuming these tools exist on some platforms, and console designs down the road may bake in similar ML assist features. That’s how upscalers have scaled previously: they start as a feature gated to new silicon and, over time, reshape expectations for visuals and performance across the ecosystem.
If you’re following the GPU wars, Redstone is an important marker. It doesn’t end the DLSS story, but it makes AMD’s case that ML-first rendering is the future of practical, high-fidelity gaming — and that control over both software and silicon is how you become relevant in that future.