AMD FSR Redstone Comes A Neural Change for Radeon RX 9000 Series
Real-time graphic rendering continues to evolve, and AMD is now really making a huge leap with its latest release of the FSR "Redstone" SDK. Breaking away from purely analytical methods, it suddenly becomes very neural rendering and machine learning (ML)-ed to power the next leap in visual fidelity. This is specifically designed to tap the architecture of the new Radeon RX 9000 (RDNA 4) series; it seeks to close the gap in AI-driven graphics performance.
Along with this, AMD streamlines its branding. The long and complex "FidelityFX Super Resolution" naming convention will be replaced by the easy-to-understand "AMD FSR." Today this umbrella term refers to all upscaling and frame generation technologies.
Redstone Pillars: Four New Technologies
There are, in fact, primarily four feature releases with the Redstone update. All machine-learning capabilities are, however, exclusive to RDNA 4 hardware. All others, which are much older than RDNA 3.5, still have analytical fallbacks.
1. FSR Upscaling (ML-Powered)
Now FSR Upscaling, formerly known as FSR 4, is based on neural net models trained on millions of very high-quality captures using AMD Instinct clusters. According to AMD, moving towards an ML-based inference model on RDNA 4 cards should enable significant improvements in the qualities of images produced, particularly regarding ghosting reduction in moving objects and better particle preservation.
Compatibility Note: In case RDNA 3.5 or older hardware is used, it automatically falls back to the analytical upscaling method (FSR 3.1 logic) to ensure that the game runs fine.
2. FSR Frame Creation
Neural treatment has been applied to frame generation as well. Using motion vectors and optical flow to interpret the current state and predict what pixels will look like in motion, the new algorithm creates superior results as it captures behavior in terms of accuracy at determining what will occur as compared to previous interpolation methods. Compared to older versions, smoother motion resulted and reduction in artifacts such as "motion discontinuities."
For the best experience, AMD recommends pairing it with the new Radeon Anti-Lag 2 SDK, which mitigates the latency in generating extra frames.
3. FSR Ray Regeneration
In modern ray reconstruction, this is by AMD. It's a denoiser in real time, on its own, and driven by ML. "Noise" here refers to rays in fewer optimums. AI is applied to clean that up and give a more accurate lighting/reflection picture without the heavy performance costs of casting more rays.
4. FSR Radiance Caching (Preview)
In looking at FSR Radiance Caching as a technical preview, AMD is moving into the future. This advanced technique leverages an online neural network to learn and estimate real-time lighting propagation for a scene in order to enable full path tracing by drastically reducing computation costs. Currently open for developer testing, it has an anticipated full production release date of 2026.
Hardware Compatibility and The RDNA 4 Split
This is what hardware is concerned in the transition to neural rendering. To be able to enjoy the feature's AI-driven versions, users should have an AMD Radeon RX 9000 series (RDNA 4) or new ones.
- Radeon RX 9000 Series: Supports ML Upscaling, ML Frame Generation, and Ray Regeneration plus Radiance Caching.
- Radeon RX 7000/6000 series (and older): Analytical fallbacks for Upscaling and Frame Generation (FSR 3.1-quality equivalent). Does not apply to the two features above: Ray Regeneration and Radiance Caching.
Availability for Gamers and Developers
The AMD FSR "Redstone" SDK will be immediately available to developers, alongside updated plugins taking Unreal Engine from version 5.2 through 5.7.
AMD has also stated that over 200 titles will support FSR Redstone upscaling and nearly 40 titles are expected to support Frame Generation by the end of 2025. Implicit or demonstrated implementations known to be coming include Call of Duty: Black Ops 6 and Warhammer 40,000: Darktide.
In addition, the new updating of the future driver updates (AMD Software: Adrenalin Edition) will allow updating the ML models used in games without the need for a game patch, bringing with it the assurance that all RDNA 4 users will always have the latest training models applied.
Source: AMD Official Announcement on Youtube and GPUOpen SDK Release Notes.
