Introduction
AMD FSR “Redstone” was first announced earlier this year at Computex as Project Redstone, and is finally available now. With Redstone, AMD is extending the existing FidelityFX Super Resolution 4 ecosystem by introducing additional machine learning–based rendering stages, focused on frame generation, ray tracing reconstruction, and lighting acceleration (Radiance Caching). These features form part of the FSR 4 feature-set AMD introduced with its new Radeon RX 9000-series GPUs powered by the RDNA 4 graphics architecture.
These features serve to close the technological gap AMD has with NVIDIA, which released DLSS 4 with its GeForce RTX 50-series. The red team made several big strides with improving the image quality of its upscaling with FSR 4, the upscaling component of which debuted alongside the RX 9000-series GPUs earlier this year. AMD implemented a new machine learning (ML) based upscaler. ML will be a recurring theme as the company extensively leveraged it for the other new components being introduced today under the banner of FSR “Redstone.”
The new features in Redstone are:
- FSR Frame Generation using a machine learning–based frame interpolation system.
- FSR Ray Regeneration, a neural network–based denoiser for improved ray tracing rendering quality.
- FSR Radiance Caching, a neural radiance cache designed to reduce the cost of ray-traced global illumination.
Given that AMD already launched FSR 4 upscaling in March, there are no changes, it’s just that the FSR 4 feature-set is being broadened to add these three new components. Radiance caching is available to developers as part of the Redstone SDK, with game integrations planned for 2026, so it’s not something we can test today, but we’ll still walk you through the technology.
In the rest of this article, we will explain the technical behavior of each Redstone component, the conditions under which they are deployed, and their impact on performance and image stability on supported hardware.
Hardware Requirements
FSR Redstone’s machine learning–based features call for GPUs based on the latest RDNA 4 graphics architecture, namely the Radeon RX 9000-series. Neural inference for frame generation, ray regeneration, and radiance caching is executed on the dedicated AI hardware units introduced with the Radeon RX 9000-series. These features are not available on earlier RDNA architectures.
This means that the Ray Regeneration and Radiance Caching aspects of FSR 4 Redstone will be exclusive to RDNA 4 hardware (Radeon RX 9000-series). The new ML-based versions of Frame Generation and the existing FSR 4 Upscaling models brought forwards require RDNA 4 hardware, too. Shader-based fallbacks exist to support these two features on older hardware, but without the improved performance and quality of the AI-based models.
| Feature | RDNA 4 | Pre-RDNA 4 |
|---|---|---|
| Upscaling | ML | FSR 3.1 Fallback |
| Frame Generation | ML | FSR 3.1 Fallback |
| Ray Regeneration | ML | No |
| Radiance Caching | ML | No |
From a game developer’s perspective, FSR Redstone features can either be implemented directly in the game, which exposes the options in the settings menu, too, or support can be delegated to the AMD Radeon Settings control panel, which takes care of overriding the game’s older FSR implementation with the newest compatible version—automatically—no manual DLL swapping required. This seems to mean FSR 4 for RX 9000-series RDNA 4 cards, and FSR 3.1 for RDNA 2 and 3 cards.